Genomic and Immune Landscape of Recurrent and/or Metastatic Squamous Cell Carcinoma of the Head and Neck Progressing on Anti–PD-1 Treatment

This multiomics characterization of R/M SCCHN progressing on anti–PD-1 therapy highlights distinct immune resistant mechanisms. MYC pathway alterations are more frequent in anti–PD-1 resistant SCCHN, while high β2M expression is associated with better overall survival.

Abstract

Anti–PD-1 therapies improve survival in recurrent/metastatic (R/M) squamous cell carcinoma of the head and neck (SCCHN), but only a minority of patients achieve durable responses. The mechanisms driving resistance to anti–PD-1 in SCCHN remain poorly understood. Using the IMMUcan multiomics workflow, we characterized the molecular and immune profiles of R/M SCCHN progressing on anti–PD-1 treatment and compared them with an anti–PD-1–naïve cohort. Tumor biopsies from patients with anti–PD-1–resistant SCCHN exhibited significantly more EGFR and MYCL amplifications, along with increased MYC pathway alterations. Transcriptomic and proteomic analyses revealed that anti–PD-1–secondary resistant SCCHN had increased CD8⁺ T-cell infiltration with higher levels of immune exhaustion markers than primary resistant and naïve SCCHN. Additionally, high beta-2-microglobulin (B2M) expression correlated with greater T-cell infiltration and improved survival following anti–PD-1 therapy. Tumor cell B2M expression was independent of TMB and PD-1L expression, suggesting that B2M expression could serve as an additional biomarker for anti–PD-1 response.

Introduction

Cancer immunotherapies targeting the programmed cell death ligand 1 (PD-L1) axis improve overall survival (OS) in many different tumor types, including recurrent and/or metastatic (R/M) squamous cell carcinoma of the head and neck (SCCHN; ref. 1). Patients with curable locally advanced SCCHN benefit from multimodal radical treatment, including chemoradiation or primary surgery followed by (chemo)radiation (2). Recently, the phase III KEYNOTE-689 trial showed that adding perioperative pembrolizumab to surgery and standard postoperative (chemo)radiation significantly improved event-free survival in resected, locally advanced SCCHN, with the greatest benefit observed in patients with a PD-L1 combined positive score (CPS) ≥1 (3); however, <60% remain cancer-free at 3 years (2). Patients with R/M SCCHN have a median survival of 10 to 17 months (2). Monotherapy with pembrolizumab as a first-line treatment improves OS in patients with tumors expressing PD-L1 (1). Similarly, first-line pembrolizumab combined with chemotherapy increases OS in R/M SCCHN regardless of PD-L1 status (1). Nivolumab and pembrolizumab also improve outcomes for patients with SCCHN after platinum therapy (4). However, only 15% to 20% of patients with R/M SCCHN treated with PD-L1 blocking monoclonal antibodies (mAb) experience prolonged survival or durable response (1, 4). These results may be partly explained by the absence of reliable predictive biomarkers of response or resistance to anti–PD-L1 therapy, as well as a lack of understanding of the mechanisms involved in immune evasion and resistance to immunotherapy.

Many studies have found predictive markers of response to PD-L1 inhibitors, such as PD-L1 expression within the tumor microenvironment (TME), abundance of tumor-infiltrating lymphocytes, tumor mutational burden (TMB), and specific immune gene signatures, but they are far from perfect, with significant overlaps between responders and nonresponders (5, 6).

In addition, the identification of immune and molecular pathways involved in resistance to immunotherapy is crucial for the development of treatment strategies to improve cancer outcomes. Several studies have suggested that genomic alterations in the antigen processing and presentation machinery (7), T-cell exhaustion (8), presence of an immune-suppressive microenvironment (suppressive cells or cytokines or chemokines; ref. 9), as well as activation of specific oncogenic pathways such as the WNT/β-catenin pathway (10) and the PI3K-AKT-mTOR pathway through the loss of PTEN (11), may be involved in resistance to immunotherapy. However, these studies primarily focused on preclinical models or cohorts of patients with cancer types other than SCCHN.

Mechanisms of resistance to immunotherapy are underexplored in SCCHN. To address this gap, we conducted a comprehensive molecular and immune profiling of R/M SCCHN samples from patients who experienced disease progression on anti–PD-1 therapy. Our study highlights several features associated with resistance to PD-1 inhibitors, providing insights that may improve patient selection and optimize immunotherapy approaches, paving the way for more personalized and effective treatments for SCCHN.

Materials and Methods

Patient cohorts and inclusion criteria

Patients were either enrolled in the EORTC-SPECTA protocol (NCT02834884), an academic translational research infrastructure for biomaterial collection, or in the EORTC-HNCG-1559 trial (NCT03088059), a biomarker-driven study that included patients with R/M SCCHN after tumor biopsy (12). Both trials are approved by the ethics committees of each participating institution and by the health authorities of each participating country and are conducted in accordance with the Declaration of Helsinki (October 2000). Written informed consent was obtained from each patient, including participation in optional translational research and genetic testing. Patients selected for this work had to meet the following inclusion criteria: (i) R/M squamous cell carcinoma of the oral cavity, oropharynx, hypopharynx, or larynx; (ii) either naïve to anti–PD-L1 treatment at the time of tumor biopsy sampling or having progressed on/after anti–PD-L1 therapy, with a tumor biopsy performed within 3 months after anti–PD-L1 discontinuation and no other treatment administered between anti–PD-L1 discontinuation and biopsy sampling; (iii) for whom whole blood and tumor biopsy performed for R/M disease were available; and (iv) for whom the clinical characteristics and disease outcomes were prospectively recorded. Recurrent SCCHN was defined according to Odense–Birmingham: same anatomic subsite or adjacent subsite within 3 cm of the primary lesion, time interval no more than 3 years, and same p16 status for oropharyngeal carcinomas (13). Second primary SCCHN were excluded, as well as locally advanced SCCHN, salivary gland, and sino-nasal cancers. The number of screened patients and reasons for exclusions are detailed in supplementary Fig. S1A.

The analyzed cohort included 126 males and 27 females, with a median age of 62 years (range 21–87). Patients originated from eight different European countries: Belgium, France, Italy, Lithuania, the Netherlands, Portugal, Spain, and the United Kingdom. The STROBE guidelines were applied.

Sample collection and processing

For each patient, a whole blood sample (EDTA vacutainer) and a fresh tumor formalin-fixed, paraffin-embedded (FFPE) tissue sample performed at the time of R/M disease were mandatory. A frozen tumor biopsy performed at the same time point as the FFPE tumor sample was optional. Samples were analyzed through the IMMUcan pipeline (14). The quality of the FFPE material was assessed centrally by a pathologist, and only samples of sufficient quality (more than 10% viable tumor cells) were analyzed.

Germline DNA (gDNA) extracted from whole blood was sequenced for all patients. Depending on the quality and availability of the samples, whole-exome sequencing (WES) and RNA sequencing (RNA-seq) were performed on FFPE tumor biopsies or on frozen tumor biopsies (Supplementary Fig. S1A). After quality control, a total of 153 patients with WES datasets were included, of which 132 had RNA-seq. Three multiplex immunofluorescence (mIF) panels and one imaging mass cytometry (IMC) panel were performed on the corresponding FFPE tumor biopsies of the 153 patients whose WES datasets passed quality control. After quality checks, mIF data from panel 1 (cell types), panel 2 (cell state), and panel 3 (T-cell subtypes; Supplementary Table S1), and IMC (15) data were available for 129, 124, 129, and 104 patients, respectively, and are included in our analyses (Fig. 1; Supplementary Fig. S1B).

Figure 1.

Study workflow. Schematic representation of the cohorts, subgroups, and analyses performed. Gray vertical boxes represent individual patients for whom the corresponding type of data is available. mIF1, mIF panel 1; mIF2, mIF panel 2; mIF3, mIF panel 3; n, number; secondary R, secondary resistant patients.

Clinical data

A patient was defined as a smoker and/or drinker if the patient had smoking history and/or a drinking history of 3 or more units of alcohol per day for a man and 2 or more units of alcohol per day for a woman. A patient was defined as a nonsmoker/nondrinker  if the patient had never smoked and/or had light consumption of alcohol (occasional or less than 3 units of alcohol per day for a man and less than 2 units of alcohol per day for women).

Primary resistance was defined as patients who progressed within 6 months after starting PD-L1 inhibitors, and secondary resistance was defined as patients who progressed more than 6 months after starting PD-L1 inhibitors. OS was defined as the time from anti–PD-L1 therapy initiation to death from any cause. For anti–PD-1–naïve patients not treated with anti–PD-L1, OS was defined as the time from biopsy sampling to death from any cause. Progression-free survival (PFS) was defined as the time from anti–PD-L1 therapy initiation to observed progression (clinically or radiographically) or death from any cause, whichever occurred first. Best overall response (BOR) was defined as the best response recorded from the start of anti–PD-L1 treatment until disease progression or the end of treatment. Clinical benefit was defined as a BOR of complete response, partial response, or stable disease lasting more than 6 months, whereas no clinical benefit was defined as a BOR of disease progression or stable disease lasting less than 6 months. Response rate was defined as the percentage of patients who have a partial response or a complete response as BOR.

Human papillomavirus status and human papillomavirus identification

The human papillomavirus (HPV) status was primarily determined by RNA-seq. To detect HPV presence in the tumor samples from bulk RNA-seq, we adapted the pipeline designed by Khan and colleagues (16). We extracted the reads that did not map to the human genome (using HISAT2 with default parameters, RRID: SCR_015530) and aligned them using Bowtie2 (RRID: SCR_016368) to a viral reference created with the genomes of all the different HPV types and the Epstein–Barr virus genome. We considered a hit if there were more than 1,000 reads concordantly and uniquely aligned to one of the viral genomes.

For patients without RNA-seq data available (n = 21), HPV-positive patients were defined as those with p16-positive oropharyngeal cancer for whom HPV was detected either by DNA polymerase chain reaction (PCR) or in situ hybridization (ISH). In total, HPV status was determined as positive for 21 patients: 20 by RNA-seq and 1 by ISH.

Genomics and transcriptomics profiling experiments

DNA and RNA from FFPE and frozen tissue were isolated using the AllPrep DNA/RNA FFPE kit (QIAGEN) or the MagMAX FFPE DNA/RNA Ultra Kit (Thermo Fisher Scientific). gDNA from EDTA tubes was isolated using the PAXgene Blood DNA kit (QIAGEN). Exome libraries were generated using the Twist Human Core Exome Kit supplemented with the RefSeq and Mitochondrial Panel for enrichment (Twist Bioscience) or the Kapa HyperExome kit (Roche). Tumor DNA was sequenced at a coverage of 100 to 150× or higher; gDNA was sequenced at a coverage of 30× or higher.

Depending on the quality and quantity of the available RNA input material, the total RNA libraries were prepared using the KAPA RNA HyperPrep Kit with RiboErase (HMR) Globin (Roche) or the SMART-Seq Stranded Kit (Takara Bio).

Next-generation sequencing was performed on an Illumina NovaSeq 6000 platform (Illumina) with a 2 × 100 bp read mode. The generated sequencing data were demultiplexed using Illumina bcl2fastq (version 2.20, RRID: SCR_015058), and adapters were trimmed with Skewer (version 0.2.2, RRID: SCR_001151).

Somatic mutation calling

WES data were analyzed using the Nextflow VEGAN pipeline (version 2.1.1) available at https://github.com/bioinfo-pf-curie/vegan. Briefly, sequencing reads were aligned onto the human hg38 genome using Bwa-mem (RRID: SCR_022192). Reads intersecting the exome capture with a minimum mapping quality of 20 were kept for downstream analysis. Duplicates were removed using the Picard MarkDuplicates tool (RRID: SCR_006525). Somatic mutations were called using paired tumor–normal samples, following GATK good practices with the Mutect2 software (RRID: SCR_026692).

Somatic variant annotation was performed with snpEff (RRID: SCR_005191). Only somatic variants with a variant allele frequency >5%, a sequencing depth of 20, and annotated as coding, nonsynonymous variants were used. Polymorphisms (based on GnomAD and 1000 Genomes Projects) were discarded.

Variants in oncogenes and tumor suppressor genes (TSG) listed in the OncoKB Cancer Gene list (downloaded on January 26, 2024, RRID: SCR_014782) or COSMIC Cancer Gene Census list (downloaded on January 26, 2024, RRID: SCR_002260) were analyzed. For oncogenes, in-frame insertions or deletions known in Cancer Hotspots and missense variants known in Cancer Hotspots were retained, as well as variants in the promoter region of TERT (c.-146, c.-124, c.-57, c.-138, c.-139). For TSG, frameshift insertions or deletions, splicing variants, stop-gained variants with nonsense-mediated decay, in-frame insertions or deletions known in Cancer Hotspots, and missense variants known in Cancer Hotspots were retained. In addition to those variants in listed oncogenes and TSG, variants in a selection of immune-related genes were included (Supplementary Table S2).

Tumor mutational burden

TMB was called using the pyTMB package included in the VEGAN pipeline. TMB high was defined as TMB ≥10 mutations/Mb; TMB low was defined as TMB <10 mutations/Mb.

Copy-number alterations analysis

Copy-number analysis was performed using the FACETS tool (RRID: SCR_026264) in matched normal–tumor mode, included in the VEGAN pipeline. For each alteration, FACETS returns the copy number and allelic status.

Amplifications were defined as a minimum of 5 copies in addition to the ploidy; deletions were defined as homozygous deletions (0 copies). Focal amplifications (≤10 Mb) of oncogenes and deletions of TSGs listed in the OncoKB Cancer Gene List (downloaded on January 26, 2024) or the COSMIC Cancer Gene Census list (downloaded on January 26, 2024) were analyzed, as well as focal amplifications and deletions in a selection of immune-related genes (Supplementary Table S2). We added E2F1 to the list of oncogenes, with E2F1 amplifications being one of the main copy-number alterations (CNA) found in HPV-positive SCCHN. Genes significantly more frequently altered in FFPE samples than in frozen samples were filtered out (Fisher exact test, P value <0.05). Moreover, only genes in which amplifications translated into higher RNA expression levels compared with the RNA levels of tumors with neutral copy numbers [Wilcoxon test, adjusted P value (Benjamini-Hochberg) < 0.05] were kept, as well as genes in which homozygous deletions translated into lower RNA expression levels compared with RNA expression levels of tumors with neutral copy numbers (Wilcoxon test, adjusted P value <0.05). Oncogenes CCND1, FGF3, FGF4, and FGF19 listed in OncoKB are located on the same 11q13.3 locus. This locus contains 10 genes mainly amplified together in SCCHN. We therefore analyzed the frequency of the locus, rather than each gene individually. Loss of heterozygosity (LOH) with neutral copy number or with a loss (defined as a number of copies below ploidy) was assessed for HLA-A, HLA-B, HLA-C, beta-2-microglobulin (B2M), and FAT1 genes. Table S3 summarizes the list of CNAs analyzed.

Genomic metrics and pathway analysis

The tumor ploidy and purity were estimated by FACETS. Patients without purity (not applicable, given by FACETS) were excluded from all analyses. The ploidy score was only reported for frozen tumor biopsies (n = 53).

To infer the fraction of the genome altered, we implemented the script infer_pga.py from FACETS with default parameters. The fraction of the genome altered (FGA) was only reported for frozen tumor biopsies (n = 53).

Fourteen selected oncogenic pathways were analyzed (17–20). For each pathway, genes common to these genes listed in the reference articles and to the oncogenes or TSGs listed in the OncoKB Cancer Gene List (downloaded on January 26, 2024) or COSMIC Cancer Gene Census list (downloaded on January 26, 2024) were analyzed (Supplementary Table S4). Per sample, a pathway was considered altered if any of the genes harbored a somatic mutation, an amplification, or a deletion as described above.

Transcriptomics data processing

Paired-end sequencing reads were aligned using STAR (version 2.7.10b, RRID: SCR_004463) against the GRCh38/hg38 human reference genome. Gene-level counts were computed with LiBiNorm using the GENCODE version 29 assembly. Whenever necessary, TPMs were computed using gene lengths from GENCODE.

Prior to the different feature extraction methods, gene expression was normalized using the Variance Stabilizing Transformation from the DESeq2 library (RRID: SCR_015687).

We applied the consensusTME method (21) on the normalized expression matrix, with the “HNSC” and “ssgsea” options.

We computed pathway activity using PROGENy (GitHub, saezlab/progeny) and cytokine activity scores using CytoSig (GitHub - data2intelligence/CytoSig).

We gathered signatures for cancer-associated fibroblasts (CAF) and epithelial–mesenchymal transition (EMT) from the literature (22–25), as well as signatures involved in response and resistance to anti–PD-L1 (bioRxiv 2015.09.01.025908; refs. 9, 26–29), and computed a score using ssgsea.

mIF analysis

Three panels of mIF were performed on consecutive slides of FFPE tumor biopsies. The stained slides were scanned using the PhenoImager HT (Akoya) with the MOTiF mode, allowing whole slide multispectral image acquisition. The scanner outputs a multichannel 8 bit image in QPTIFF format. The first panel was designed to detect immune cell populations. It contains antibodies against CD15 (cat. #301902; RRID:AB_314194), a phenotypic marker of neutrophils; CK (cat. #M3515; RRID:AB_2132885), a tumor marker; CD163 (cat. #MB460; RRID:AB_3714914), a phenotypic marker of macrophages; CD11c (cat. #111M-15; RRID:AB_3714915), a phenotypic marker of dendritic cells (DC); CD20 (cat. #M0755; RRID:AB_2282030), a phenotypic marker of B cells; and CD3 (cat. #A0452; RRID:AB_2335677), a phenotypic marker of T lymphocytes. The second panel was designed to identify the cell state of tumor cells and CD8⁺ T cells. It contains antibodies against CD8 (cat. #108R; RRID:AB_2892088), CK, PD-1 (cat. #315M-96; RRID:AB_1160829), PD-L1 (cat. # 13684; RRID:AB_2687655), Ki67 (cat. # 275R-16; RRID:AB_1158037), and granzyme B (GB; cat. #MON7029C; RRID:AB_2114692). The third panel was designed to identify the different T cell subtypes. It contains antibodies against CD3, CD4 (cat. #104R; RRID:AB_1516770), CD8, FoxP3 (cat. #ab99963; RRID:AB_10675258), CD56 (cat. # 156R-9; RRID:AB_2941091), and CK. Antibody references are available in Table S1. Staining procedures are described in details by Eling and colleagues (15).

mIF image analysis was performed with the IFQuant software, as previously described (15). Briefly, tissue segmentation was performed using the tumor cell marker (CK) to assign each cell to the tumor tissue region or to the stroma region. With panel 1, we identified tumor cells and six different immune cell phenotypes: B cells, T cells, B near T cells (cells positive for CD20 and CD3), neutrophils, DCs, and macrophages CD163⁺ (MacCD163). Cells not assigned to any of the phenotypes were defined as “other.”

With panel 3, we identified six different T-cell and natural killer (NK) cell phenotypes: CD4⁺ T cells, CD8⁺ T cells, CD4⁺CD8⁺ T cells, regulatory T cells (Treg), NK cells, and NK T cells. Marker combinations for each cell type are available in Supplementary Table S5.

With panel 2, we identified and analyzed two tumor phenotypes: tumor cells Ki67⁺ (CK⁺Ki67⁺) and tumor cells PD-L1⁺ (CK⁺PD-L1⁺), and three CD8 phenotypes: CD8 Ki67⁺, CD8 PD-1⁺, and CD8 GB⁺.

Proportions of each immune cell phenotype were computed for the total region (dividing by the total number of cells), for the stroma region (dividing by the total number of cells present in the stroma region), and for the tumor region (dividing by the total number of cells present in the tumor region).

Proportions of each tumor phenotype were calculated by the sum of tumor cells positive for the marker divided by the total number of tumor cells.

Immune phenotypes were defined based on the proportion of CD8⁺ T cells present in the different regions. A tumor sample was defined as CD8⁺ T cells immune infiltrated if the sum of the proportions of CD8⁺ T cells and CD4⁺CD8⁺ T cells in the tumor region was higher than 3%, CD8⁺ T cells immune excluded if the sum of the proportions of CD8⁺ T cells and CD4⁺CD8⁺ T cells in the stroma region was higher than 3%, and CD8⁺ T cells immune desert if the sum of the proportions of CD8⁺ T cells and CD4⁺CD8⁺ T cells in the tumor and stroma regions was lower than 3%.

Tertiary lymphoid structures (TLS) were defined as patches of B cells (CD20⁺ cells), with a local B-cell density above 2,000 cells/mm² and at least 40 cells (15). TLS were considered present in a sample if at least one TLS was identified in the region of interest.

IMC analysis

IMC was performed on a consecutive slide of the FFPE tumor biopsy. Staining procedures and data processing have been previously described (15).

Data processing has been previously described (15). To assess cell phenotypes, a binarization of marker intensity was performed thanks to a Gaussian mixture model. For this work, the density of 35 different cell phenotypes was identified and analyzed (Supplementary Table S6). Per sample, three regions of interest were stained. ROIs were selected based on the IF slide, focusing on regions containing tumor and immune cells (15). The mean density of each tumor cell phenotype from the three regions of interest was computed per sample. For the cell phenotype “tumor cells B2M⁺,” we also computed the proportion of tumor cells B2M⁺ divided by the total number of tumor cells. For the binarization of this variable, we used as a cutoff the inflection point of the B2M⁺ distribution in the whole cohort (20%).

Data quality control

We verified the matching of samples among gDNA, tumor DNA, and tumor RNA, and we confirmed the correspondence between the clinical gender variable and RNA gender determination.

The verification that the gDNA, tumor DNA, and tumor RNA fastq files originated from the same donor was performed using an in silico genotyping test. A custom Perl script first extracted the allele frequency at each genomic position specified in the bed file associated with the HumanCore-12 v1-0 Illumina kit and then compared the overall concordance frequency of all alleles. For two patients, gDNA did not correspond to tumor DNA and RNA. Those patients were discarded.

We determined the transcriptomic sex and compared it with clinical sex as a rapid quality measure by looking at the expression of XIST and DDX3Y (female and male specific genes, respectively). Following this procedure, two samples were identified for potential gender mislabeling. These data were later verified with the patients’ site. Both were confirmed by the sites as a data entry error and were corrected in the final clinical data version.

For 100 patients, we analyzed WES datasets from FFPE tumor biopsies, and for 53 patients, WES datasets from frozen tumor biopsies (Supplementary Fig. S1B). No significant difference in TMB and single-nucleotide variants was observed between FFPE and frozen tumor biopsies. For CNA, significant differences were observed between FFPE and frozen tumor biopsies: FFPE tumor biopsies harbored significantly more deletions and had higher ploidy and higher FGA (Supplementary Fig. S1C). In addition to those tests, we analyzed 12 patients’ paired WES datasets coming from a frozen tumor biopsy and the corresponding FFPE tumor biopsy. Paired tests were performed, and the same results were obtained (Supplementary Fig. S1D). We, therefore, focused our copy-number analysis on genes that were not significantly more frequently altered in FFPE tumor biopsies than in frozen tumor biopsies, and we report the ploidy and FGA only for patients with frozen tumor biopsies analyzed (n = 53).

For RNA-seq analyses, 81 patients had RNA-seq datasets from FFPE tumor biopsies, and 51 patients had RNA-seq datasets from frozen tumor biopsies. The variable sample conservation type (FFPE or frozen tumor biopsy) was always included in the multivariate analyses.

mIF and IMC analyses were always performed on the FFPE tumor biopsies.

Quantification and statistical analysis

Statistical analyses were performed using R version 4.1.0. For frequencies, a Fisher’s exact test was used. For continuous variables with two categories, a Wilcoxon test was used; for continuous variables with more than two categories, a Kruskal–Wallis test was performed. Correlation tests were performed using the Spearman method. When mentioned, corrections for multiple testing were performed using the Benjamini-Hochberg (BH) method. Multivariate analyses were performed using a linear regression model.

To assess the differences in TMB, a linear regression model was fitted, including technical variables [DNA assay library, sample conservation type (FFPE vs. frozen), and biopsy sampling site] and clinical variables (HPV status and number of lines of systemic treatment administered for R/M disease). For the FGA, a linear regression model was fitted, including technical variables (DNA assay library and biopsy sampling site) and the same clinical variables.

To find differences in transcriptomic features according to our different groups, we fitted a linear model including as covariates the sample conservation type (FFPE or frozen), the RNA extraction method, the RNA library protocol used, and the sample size (in a binarized form, taking the median as the threshold). We corrected the obtained P values using the BH correction.

OS and PFS were estimated using the Kaplan–Meier method. Between-group differences in OS and PFS were assessed using the stratified log-rank test. A stratified Cox proportional hazards model was used to estimate HRs and their associated 95% confidence intervals (CI), as well as to perform multivariate analyses incorporating key prognostic variables.

Additional resources

Additional information is contained in these studies: NCT03088059: EORTC-HNCG-1559 and NCT02834884: EORTC-SPECTA protocol.

Results

Patient cohorts and multiomics analyses

We aimed to identify immune and molecular mechanisms implicated in resistance to PD-L1 inhibitors in R/M SCCHN. Two cohorts were selected: 74 patients progressed on PD-1 (n = 73) or PD-L1 (n = 1) inhibitors at the time of tumor biopsy (anti–PD-1–resistant cohort), and 79 patients were naïve to any anti–PD-L1 treatment at biopsy (anti–PD-1–naïve cohort). The cohorts were clinically well balanced, except for treatment history, with anti–PD-1–resistant patients having received more therapy lines in the R/M setting (Table 1). Primary diagnosis sites were the oral cavity for 36 (24%) patients, oropharynx for 67 (44%), hypopharynx for 28 (18%), and larynx for 22 (14%). Twenty-one (14%) patients were HPV-positive.

Table 1.

Clinical characteristics of anti–PD-1–naïve and anti–PD-1–resistant cohorts.

​	Anti–PD-1–naïve cohort (N = 79)	Anti–PD-1–resistant cohort (N = 74)
		Total (N = 74)	P value (naïve vs. resistant)a	Primary resistant (N = 60)	Secondary resistant (N = 14)	P value (primary vs. secondary)a
Sex	​	​	​	​	​	​
Female	18 (22.8%)	9 (12.2%)	0.131	5 (8.3%)	4 (28.6%)	0.103
Male	61 (77.2%)	65 (87.8%)	​	55 (91.7%)	10 (71.4%)	​
Age (years)	​	​	​	​	​	​
Median (min, max)	61 (21, 81)	62 (36, 87)	0.101	62 (40, 87)	63 (36, 75)	0.741
ECOG
0	17 (21.5%)	21 (28.4%)	0.199	18 (30%)	3 (21.4%)	0.608
1	49 (62%)	51 (68.9%)	​	40 (66.7%)	11 (78.6%)	​
2	7 (8.9%)	2 (2.7%)	​	2 (3.3%)	0 (0%)	​
Missing	6 (7.6%)	0 (0%)	​	0 (0%)	0 (0%)	​
Substance abuse
Nonsmoker and nondrinker	11 (13.9%)	12 (16.2%)	0.865	10 (16.7%)	2 (14.3%)	1
Smoker and/or drinker	68 (86.1%)	62 (83.8%)	​	50 (83.3%)	12 (85.7%)	​
HPV status
HPV negative	68 (86.1%)	64 (86.5%)	1	51 (85%)	13 (92.9%)	0.734
HPV positive	11 (13.9%)	10 (13.5%)	​	9 (15%)	1 (7.1%)	​
Primary disease location
Oral cavity	19 (24.1%)	17 (23%)	0.986	16 (26.7%)	1 (7.1%)	0.415
Oropharynx	34 (43%)	33 (44.6%)	​	26 (43.3%)	7 (50%)	​
Hypopharynx	14 (17.7%)	14 (18.9%)	​	11 (18.3%)	3 (21.4%)	​
Larynx	12 (15.2%)	10 (13.5%)	​	7 (11.7%)	3 (21.4%)	​
Primary disease stage
I–II	12 (15.2%)	9 (12.2%)	0.081	8 (13.3%)	1 (7.1%)	0.103
III	9 (11.4%)	9 (12.2%)	​	8 (13.3%)	1 (7.1%)	​
IVA-B	31 (39.2%)	39 (52.7%)	​	33 (55%)	6 (42.9%)	​
IVC	21 (26.6%)	8 (10.8%)	​	4 (6.7%)	4 (28.6%)	​
Missing	6 (7.6%)	9 (12.2%)	​	7 (11.7%)	2 (14.3%)	​
Primary disease grade
Grade 1	14 (17.7%)	17 (23%)	0.852	15 (25%)	2 (14.3%)	0.554
Grade 2	20 (25.3%)	28 (37.8%)	​	21 (35%)	7 (50%)	​
Grade 3	13 (16.5%)	21 (28.4%)	​	17 (28.3%)	4 (28.6%)	​
Missing	32 (40.5%)	8 (10.8%)	​	7 (11.7%)	1 (7.1%)	​
Disease extent at the time of tumor biopsy
Locoregional only disease	32 (40.5%)	26 (35.1%)	0.605	22 (36.7%)	4 (28.6%)	0.795
Distant metastatic disease	47 (59.5%)	48 (64.9%)	​	38 (63.3%)	10 (71.4%)	​
Previous curative intent surgery
No	41 (51.9%)	39 (52.7%)	1	29 (48.3%)	10 (71.4%)	0.207
Yes	38 (48.1%)	35 (47.3%)	​	31 (51.7%)	4 (28.6%)	​
Previous curative intent radiotherapy
No	30 (38%)	19 (25.7%)	0.145	10 (16.7%)	9 (64.3%)	<0.001
Yes	49 (62%)	55 (74.3%)	​	50 (83.3%)	5 (35.7%)	​
Number of R/M treatment lines prior to biopsy
0	57 (72.2%)	5 (6.8%)	<0.001	4 (6.7%)	1 (7.1%)	0.839
1	17 (21.5%)	23 (31.1%)	​	18 (30%)	5 (35.7%)	​
2	4 (5.1%)	35 (47.3%)	​	28 (46.7%)	7 (50%)	​
3+	1 (1.3%)	11 (14.9%)	​	10 (16.7%)	1 (7.1%)	​
Anti–PD-L1 regimen before tumor biopsy
ICI alone	NA	62 (83.8%)	​	52 (86.6%)	10 (71.4%)	0.322
ICI with chemotherapy	NA	12 (16.2%)	​	8 (13.4%)	4 (28.6%)	​
Anti–PD-L1 regimen after tumor biopsy
ICI alone	32 (40.5%)	NA	​	NA	NA	​
ICI with chemotherapy	16 (20.3%)	NA	​	NA	NA	​
Anti–PD-L1 treatment duration (days)
Median (min, max)	NA	102 (30, 797)	NA	91 (30, 192)	279 (189, 797)	<0.001
Anti–PD-L1 BORb
Complete response	3 (3.8%)	0 (0%)	NA	0 (0%)	0 (0%)	<0.001
Partial response	6 (7.6%)	7 (9.5%)	​	1 (1.7%)	6 (42.9%)	​
Stable disease	11 (13.9%)	26 (35.1%)	​	18 (30%)	8 (57.1%)	​
Progressive disease	17 (21.5%)	41 (55.4%)	​	41 (68.3%)	0 (0%)	​
Missing	11 (13.9%)	0 (0%)	​	0 (0%)	0 (0%)	​
NA	31 (39.2%)	​	​	​	​	​

Abbreviations: ECOG PS, Eastern Cooperative Oncology Group performance status; ICI, immune checkpoint inhibitor; max, maximum; min, minimum; NA, nonapplicable.

^a P value derived from Fisher's exact test.

^bBOR corresponds, in the anti–PD-1–naïve cohort, to patients who received anti–PD-1 therapy after tumor biopsy and, in the anti–PD-1–resistant cohort, to the best response achieved with anti–PD-1 prior to the development of resistance.

Among the 74 resistant patients, 60 had primary resistance (progression within 6 months of anti–PD-L1 initiation), and 14 had secondary resistance (progression more than 6 months after treatment start). Median anti–PD-L1 treatment duration was 102 days (range: 30–797 days). Clinical characteristics of anti–PD-1–primary resistant and anti–PD-1–secondary resistant patients are summarized in Table 1. Fewer secondary resistant patients had received curative-intent radiotherapy, including lymph node irradiation, compared with primary resistant patients (36% vs 83%, nominal P < 0.001, Fisher’s exact test). Tumor biopsies were obtained within 3 months after anti–PD-L1 discontinuation (median 31 days, range: 0–81 days). No anticancer treatments were administered between the end of the anti–PD-L1 therapy and the tumor biopsy collection.

Among the 79 anti–PD-1–naïve patients, 48 (61%) were subsequently treated with anti–PD-L1 therapy, with or without chemotherapy (n = 16 and 32, respectively).

To characterize anti–PD-1 resistance in R/M SCCHN at an unprecedented scale, tumor biopsies were analyzed using the IMMUcan workflow, combining genomic, transcriptomic, and spatial cellular profiling. WES was available for all 153 patients, whereas transcriptomic datasets were available for 132 (Fig. 1). We analyzed a set of 96 transcriptomic features, including deconvolution scores (21) for immune cells and a few other cell types, activation scores for several cancer-related and cytokine pathways (see “Materials and Methods”), EMT signatures (23, 24), CAF signatures (22, 25), and six signatures associated with response (bioRxiv 2015.09.01.025908; refs. 26, 27) or resistance (9, 28, 29) to immunotherapy. The spatial cellular profiling was generated using mIF staining and IMC (15) to investigate T cell and other immune or nonimmune cell types and states. Data from the three mIF panels and the IMC panel were available for 129, 124, 129, and 104 patients, respectively.

EGFR, MYCL, and RRAGC amplifications are more frequent in anti–PD-1–resistant SCCHN compared with anti–PD-1–naïve SCCHN

To identify genomic alterations that could be associated with resistance to PD-L1 inhibitors in SCCHN, we identified the genes that were more frequently altered in the anti–PD-1–resistant cohort compared with the anti–PD-1–naïve cohort.

The most frequent genomic alterations in the anti–PD-1–resistant and anti–PD-1–naïve cohorts are depicted in Supplementary Fig. S2A. Figure 2 shows the genomic alterations found at a higher frequency in the anti–PD-1–resistant cohort compared with the anti–PD-1–naïve cohort. PIK3CA (16% vs. 5%, nominal P < 0.05, Fisher’s exact test) and CREBBP (7% vs. 0%, P < 0.05) mutations and EGFR (14% vs. 1%, P < 0.01), MYCL, and RRAGC (8% vs. 0%, P < 0.05) amplifications were significantly more common in anti–PD-1–resistant SCCHN (Fig. 2). Of note, MYCL and RRAGC are both located on locus 1p34.2. EGFR, MYCL, and RRAGC amplifications remained significant (P < 0.05) after adjusting for HPV status and prior treatments. No differences in the FGA, which is the fraction of the genome encompassed by CNAs, or TMB were found between the two cohorts after correcting for the number of R/M systemic treatment lines, HPV status, and technical variables (Supplementary Fig. S2B).

Figure 2.

Genomic differences between anti–PD-1–naïve and anti–PD-1–resistant R/M SCCHN cohorts. OncoPrint of genes more frequently mutated in the anti–PD-1–resistant cohort compared with the anti–PD-1–naïve cohort (for visualization purposes, a P value cutoff of P < 0.15 was selected), as well as a selection of genes that were previously shown (7,11,30) to be potentially involved in resistance mechanisms to anti–PD-L1 therapy (n = 153). For the patients subsequently treated with anti–PD-L1 therapy, clinical benefit was defined as a BOR of complete response, partial response, or stable disease lasting more than 6 months, whereas no clinical benefit was defined as a BOR of disease progression or stable disease lasting less than 6 months. NMD, nonsense-mediated decay.

We examined genes implicated in immune evasion or anti–PD-1 resistance (Supplementary Table S7; refs. 7, 11) for an association with the anti–PD-1–resistant cohort. Only a few were sparsely altered in the anti–PD-1–resistant cohort (Fig. 2): mutations in STK11 (n = 1), IDO1 (n = 3), PSMB10 (n = 1), B2M (n = 1), or homozygous deletions of PTEN (n = 2) and B2M (n = 1) were found in resistant patients only. TAPBPL, a negative regulator of T-cell function, was amplified in 4% of anti–PD-1–resistant samples, but only in 1% of anti–PD-1–naïve samples (Fig. 2). Genes associated with antigen presentation (TAP1, ERAP1, ERAP2, NLRC5) or the IFNγ pathway (STAT1, JAK2, and IRF1) were rarely mutated (<2%) and were not enriched in the anti–PD-1–resistant cohort (Fig. 2). LOH in at least one HLA class I gene occurred equally (46%) in both cohorts, and B2M LOH was also similar (15% and 24% in anti–PD-1–naive or anti–PD-1–resistant, respectively; nominal P = 0.30, Fisher’s exact test). Thus, no major differences were observed in immune evasion gene alterations between anti–PD-1–resistant and anti–PD-1–naïve SCCHN.

Finally, we examined whether some oncogenic pathways were differently activated between our two cohorts (Supplementary Table S4). The MYC pathway was altered in tumor biopsies of 13 (18%) of the 74 anti–PD-1–resistant patients and only in one (1%) of the anti–PD-1–naïve patients (nominal P < 0.001, Fisher’s exact test; multivariate P < 0.01, linear model; Fig. 2). No differences in the other pathways were observed.

In summary, genomic analyses showed that EGFR amplifications and MYC pathway alterations were more frequent in anti–PD-1–resistant SCCHN compared with anti–PD-1–naïve cases.

MYC pathway and EGFR genomic alterations are associated with lower immune infiltration

We further investigated the association between MYC pathway alterations and modifications in the transcriptome and the immune microenvironment. Thirty-seven (38%) of the aforementioned transcriptomic features showed significant differences between SCCHN with or without MYC-altered pathways (adjusted P < 0.05, linear model). All significantly different features were lower in MYC-altered tumors (Fig. 3A). They included the majority of the CAF signatures, signatures related to TGFβ signaling (TGFβ score from PROGENy, TGFβ1, BMP4, and activin A scores from CytoSig), the JAK-STAT score, and deconvoluted fibroblasts, endothelial cells, and myeloid cells. Finally, a number of cytokine pathways also showed lower activation scores in MYC-altered tumors (GCSF, MCSF, CXCL12, TNFα, and IL22). Of note, B and T lymphocyte signatures were also lower in MYC-altered SCCHN (adjusted P < 0.1, linear model).

Figure 3.

MYC pathway and EGFR genomic alterations are associated with lower immune infiltration. A, Bar plot of significantly different transcriptomic features between MYC-altered (n = 14) and MYC-WT SCCHN (n = 118). The P values from linear model (see “Materials and Methods”). Dashed red line indicates the threshold of adjusted P value = 0.05 (BH method). B, Immune cell proportions in the tumor (left) and stroma regions (right) in MYC-altered (n = 12) and MYC WT SCCHN (n = 117). The P value was from the Wilcoxon test. C, Proportions of NK and T cell subtypes in the tumor (left) and stroma regions (right) in MYC altered (n = 12) and MYC-WT SCCHN (n = 117). The P value was from the Wilcoxon test. D, Immune cell proportions in the tumor (left) and stroma regions (right) in SCCHN tumors with EGFR amplification (n = 11) vs. EGFR WT (n = 118). The P value was from the Wilcoxon test. E, Proportions of NK and T cell subtypes in the tumor (left) and stroma regions (right) in SCCHN tumors with EGFR amplification (n = 10) vs. EGFR WT (n = 119). The P value was from the Wilcoxon test. Boxplot lines represent the 25%, 50%, and 75% percentiles. BnT cell, B near T cell defined as CD20⁺ and CD3⁺ cells; HN CAF sig, head and neck CAF signature; MacCD163, macrophages CD163⁺; NKT, natural killer T cells; ns, not significant; TGFb, TGFβ; TNFa, TNFα.

In line with these results, mIF analysis also showed lower proportions of the different immune populations in MYC-altered SCCHN, with the exception of CD163⁺ macrophages (Fig. 3B and C). These differences were observed mostly in the tumor regions and were statistically significant for CD4⁺ T cells (nominal P < 0.05, Wilcoxon rank-sum test) and Tregs (nominal P < 0.01, Wilcoxon rank-sum test).

As EGFR amplification was also significantly more frequent in the anti–PD-1–resistant cohort compared with the anti–PD-1–naïve cohort, we investigated modifications in the TME associated with this alteration. The transcriptomic analysis did not reveal significant differences, but mIF showed that tumor samples with EGFR amplification were characterized by less DCs in the tumor and stroma regions and less B cells in the stroma region (nominal P < 0.05, Wilcoxon rank-sum test; Fig. 3D and E) compared with the rest of the samples.

These data, therefore, indicate that somatic genomic alterations in SCCHN are associated with distinct cellular changes in the TME.

B2M LOH is more frequent in anti–PD-1–secondary resistant compared with anti–PD-1–naïve or anti–PD-1–primary resistant SCCHN, but this is not associated with reduced B2M expression on tumor cells

Given that primary and secondary resistance mechanisms may differ, we explored whether there were differences in gene alterations between primary and secondary resistant SCCHN. B2M LOH was significantly more frequent in secondary resistant compared with primary resistant (50% vs. 17%, nominal P < 0.01, Fisher’s exact test) and anti–PD-1–naïve SCCHN (50% vs. 15%, nominal P < 0.05, Fisher’s exact test; Supplementary Fig. S2C). Of note, mutations in H3C12, which codes for a histone, and in LRIG3, which could play a role as a negative regulator of EGFR signaling (31), were each observed in two secondary resistant SCCHN but not in the primary resistant SCCHN (nominal P < 0.05, Fisher’s exact test). TMB and FGA were not significantly different between the primary and secondary resistant groups after correcting for the number of R/M systemic treatment lines, HPV status, and technical variables (Supplementary Fig. S2D).

B2M⁺ tumor cells were more abundant in secondary resistant patients than in primary resistant or anti–PD-1–naïve patients (nominal P < 0.05 and P < 0.01, respectively, Fig. 4A). As alterations in the MHC class I have been implicated in resistance to anti–PD-1 (30), we further investigated B2M with the other omics approaches. B2M genomic alterations did not correlate with reduced B2M expression levels measured with RNA-seq (Supplementary Fig. S3A). Therefore, we assessed the effect of B2M genomic alterations on B2M protein expression in SCCHN biopsies for which we had WES and IMC data available (n = 104). B2M protein was undetectable on tumor cells by IMC in one anti–PD-1–primary resistant SCCHN with a homozygous deletion in B2M and another SCCHN with a frameshift B2M mutation in addition to LOH (Fig. 4B). We next differentiated B2M LOH with neutral copy number (LOH neutral) and B2M LOH with copy number below ploidy (LOH loss). Compared with SCCHN tumor samples with no B2M genomic alterations, SCCHN with B2M LOH loss had a lower proportion of tumor cells expressing B2M (nominal P < 0.05, Wilcoxon rank-sum test), whereas SCCHN B2M LOH neutral had no decreased proportions of tumor cells B2M⁺ (Fig. 4B). It remains unclear whether the increased frequency of B2M LOH in patients with secondary resistance has a functional impact.

Figure 4.

Some genomic alterations are associated with tumor B2M expression levels, and high tumor B2M expression is associated with increased T-cell infiltration. A, Densities of tumor cells B2M⁺ in anti–PD-1–naïve (n = 59), anti–PD-1–primary resistant (n = 37), and anti–PD-1–secondary resistant SCCHN (n = 8). The P value was from the Wilcoxon test. B, Proportions of tumor cells B2M⁺ according to B2M genomic alterations. SCCHN harboring mutations in IRF1, JAK2, and NFKB1 are highlighted in color. C, Immune cell proportions in the tumor (left) and stroma regions (right) in B2M-low (n = 54) and B2M-high SCCHN (n = 42). The P value was from the Wilcoxon test. D, Proportions of NK and T cell subtypes in the tumor (left) and stroma regions (right) in B2M-low (n = 54) and B2M-high SCCHN (n = 42). The P value was from the Wilcoxon test. Boxplot lines represent the 25%, 50%, and 75% percentiles. BnT cell, B near T cells defined as CD20⁺ and CD3⁺ cells; MacCD163, macrophages CD163⁺; NKT, natural killer T cells; SNV, single-nucleotide variant.

To note, some SCCHN with no genomic alterations in B2M also had a low proportion of B2M⁺ tumor cells, reflecting that other genomic and epigenomic mechanisms are affecting tumor cell B2M expression (30). Therefore, we investigated whether alterations in other genes implicated in antigen presentation machinery (30) could affect B2M expression. One SCCHN had an IRF1 mutation in addition to a B2M LOH loss, and none of the tumor cells were positive for B2M. Two other patients had a JAK2 mutation: one with no alteration in B2M and one with B2M LOH loss. Both had low proportions of tumor cells expressing B2M (<7%). Another patient with an NFKB1 mutation also had a low proportion of tumor cells expressing B2M (<1%; Fig. 4B). In contrast, we did not find any association between the proportion of B2M⁺ tumor cells and LOH in the heavy chain of MHC class I (HLA-A, HLA-B or HLA-C; Supplementary Figure S3B), nor with other alterations implicated in MHC class I expression (mutations in STAT1, ERAP1, ERAP2, PSMB10, NLRC5, or amplifications of JAK2; Supplementary Fig. S3C; ref. 30).

Altogether, these data suggest that B2M homozygous alterations, B2M LOH loss, and mutations in JAK2, IRF1, and NFKB1 are associated with reduced B2M protein expression on tumor cells.

High tumor B2M expression is associated with increased T-cell infiltration

We next studied the impact of B2M genomic alterations and B2M protein levels on immune cell infiltration. No difference in immune cell infiltration was observed by mIF between SCCHN with B2M LOH (neutral and loss) and B2M wild type (WT), nor between B2M LOH loss and B2M WT (Supplementary Fig. S3D and S3E). To investigate the relationship between B2M protein expression on tumor cells and immune infiltration, we defined tumors as B2M-high or B2M-low based on whether more than 20% or less than 20% of tumor cells were positive for B2M, respectively. A threshold of 20% was chosen based on the shape of its distribution in our cohort and its inflection point (Supplementary Fig. S3F). Although B2M-low tumors were more infiltrated with neutrophils in the stroma (Fig. 4C), B2M-high tumors were significantly more infiltrated with T cells in the tumor and stroma regions, namely more CD4⁺ T cells, CD4⁺CD8⁺ T cells, CD8⁺ T cells, and Tregs in the stroma (nominal P < 0.05, Wilcoxon rank-sum test) and tumor (nominal P < 0.01, Wilcoxon rank-sum test) regions, respectively (Fig. 4C and D).

These data suggest that B2M-high tumors are more infiltrated with T cells compared with B2M-low tumors.

Anti–PD-1–secondary resistant SCCHN show enhanced immune infiltration and higher expression of immune exhaustion markers compared with anti–PD-1–naïve or anti–PD-1–primary resistant SCCHN

We examined the transcriptomic datasets to better characterize anti–PD-1–resistant SCCHN. None of the 96 transcriptional features were significantly different between the anti–PD-1–naive and anti–PD-1–resistant cohorts. However, 16 of the 96 features (17%) were significantly different between anti–PD-1–primary resistant and anti–PD-1–secondary resistant patients (adjusted P < 0.05, linear model). Differentially expressed features were all related to immune response and included (i) all T-cell types and NK cells obtained by deconvolution; (ii) “T-cell exhausted,” “T cell–inflamed gene expression profile (GEP),” and “cytotoxic” signatures; (iii) immune and IFNγ scores; (iv) antigen-presenting machinery signature (APM); and (v) other immune cell types (M1 macrophages, DCs, and B cells; Fig. 5A and B). Enrichment scores of all these features were higher in anti–PD-1–secondary resistant patients, reflecting enhanced immune cell infiltration and activation within the TME in comparison with anti–PD-1–primary resistant SCCHN. This observation was supported by the comparison of secondary resistant SCCHN with the anti–PD-1–naïve SCCHN (Fig. 5A and B). In that analysis, results were similar for all of the aforementioned features, as well as for those related to other immune cell types like mast cells, neutrophils, eosinophils, and monocytes, and the interleukin 1B (IL1B) activity score (Fig. 5A and B).

Figure 5.

Anti–PD-1–secondary resistant SCCHN shows enhanced immune infiltration and higher expression of immune exhaustion markers. A, Bar plot of significantly different transcriptomic features between anti–PD-1–secondary resistant SCCHN (n = 12) compared with anti–PD-1–primary resistant SCCHN (n = 55) in red and compared with anti–PD-1–naïve SCCHN (n = 65) in blue. The P values were from linear model (see “Materials and Methods”). Dashed red line indicates the threshold of adjusted P value = 0.05 (BH method). B, Boxplots illustrating two transcriptomic features significantly different between anti–PD-1–secondary resistant SCCHN (n = 12) and anti–PD-1–primary resistant (n = 55) or anti–PD-1–naïve SCCHN (n = 65). T-cell exhausted signature (top) and Tregs (bottom). C, Immune cell proportions in the tumor (left) and stroma regions (right) in anti–PD-1–naïve (n = 68), anti–PD-1–primary resistant (n = 48), and anti–PD-1–secondary resistant SCCHN (n = 13). The P value was from the Wilcoxon test. D, Proportions of NK and T cell subtypes in the tumor (left) and stroma regions (right) in anti–PD-1–naïve (n = 66), anti–PD-1–primary resistant (n = 49), and anti–PD-1–secondary resistant SCCHN (n = 11). The P value was from the Wilcoxon test. E, Proportions of immune phenotypes in anti–PD-1–naïve (n = 69), anti–PD-1–primary resistant (n = 49), and anti–PD-1–secondary resistant SCCHN (n = 11). Fisher’s exact test P value = 0.039. F, Densities of CD163⁺ macrophages, CD40⁺, IDO1⁺, PD-L1⁺, or VISTA⁺ in anti–PD-1–naïve (n = 59), anti–PD-1–primary resistant (n = 37), and anti–PD-1–secondary resistant SCCHN (n = 8). The P value was from the Wilcoxon test. G, Densities of tumor cells IDO1⁺ or PD-L1⁺ or VISTA⁺ in anti–PD-1–naïve (n = 59), anti–PD-1–primary resistant (n = 37), and anti–PD-1–secondary resistant SCCHN (n = 8). The P value was from the Wilcoxon test. H, mRNA normalized expression of CTLA4, HAVCR2, LAG3, PDCD1, and TIGIT in anti–PD-1–naïve (n = 66), anti–PD-1–primary resistant (n = 48), and anti–PD-1–secondary resistant SCCHN (n = 13). The P value was from the Wilcoxon test. HAVCR2 codes for TIM3 and PDCD1 codes for PD-1. I, Densities of CD8⁺LAG3⁺ T cells (left), CD4⁺LAG3⁺ T cells (middle), Tregs LAG3⁺ (right) in anti–PD-1–naïve (n = 59), anti–PD-1–primary resistant (n = 37), and anti–PD-1–secondary resistant SCCHN (n = 8). The P value was from the Wilcoxon test. J, Proportion of CD8⁺LAG3⁺/total CD8⁺ T cells (left), proportion of CD4⁺LAG3⁺/total CD4⁺ T cells (middle), and proportion of Treg LAG3⁺/total Tregs (right) in anti–PD-1–naïve (n = 59), anti–PD-1–primary resistant (n = 37), and anti–PD-1–secondary resistant SCCHN (n = 8). The P value was from the Wilcoxon test. Boxplot lines represent the 25%, 50% and 75% percentiles. BnT cell, B near T cell defined as CD20⁺ and CD3⁺ cells; GEP, gene expression profile; IFNG, IFNγ; MacCD163, macrophages CD163⁺; N, anti–PD-1 naïve; NKT, natural killer T cells; ns, not significant; 1R, anti–PD-1 primary resistant; 2R, anti–PD-1 secondary resistant.

To further assess the immune contexture and to confirm our RNA-seq findings, we explored the mIF data and confirmed no differences in any of the studied immune cell types in either stroma or tumor between anti–PD-1–naïve and anti–PD-1–resistant cohorts (Supplementary Fig. S4A and S4B). Stromal B cell proportions were lower in anti–PD-1–resistant SCCHN (nominal P < 0.01, Wilcoxon rank-sum test), but this association lost significance in the multivariate analysis (Supplementary Fig. S4A). Similar results were observed for the presence of TLS (Supplementary Fig. S4C). However, in accordance with the transcriptomic data, secondary resistant SCCHN had more total T cells than anti–PD-1 primary resistant (P < 0.05; Fig. 5C) and, in particular, significantly more CD8⁺ T cells compared with anti–PD-1 primary resistant and anti–PD-1 naive in the tumor (3.7% vs. 1.3% or 1.2%, nominal P < 0.01, Wilcoxon rank-sum test; multivariate P < 0.05, linear model) and stroma regions (7.3% vs 3.4% or 3%, nominal P < 0.01, Wilcoxon rank-sum test; multivariate P < 0.05, linear model; Fig. 5D). This was in line with the computed immune phenotypes: 45% of the secondary resistant SCCHN had a T-cell CD8⁺-infiltrated phenotype and 55% had a T-cell CD8⁺-excluded phenotype, whereas none had a T-cell CD8⁺-desert phenotype. In contrast, 36% of primary resistant and anti–PD-1–naïve SCCHN had a CD8⁺-desert phenotype (Fig. 5E; nominal P < 0.05, Wilcoxon rank-sum test). Secondary resistant SCCHN had more Tregs in the stroma region compared with the anti–PD-1–naïve cohort (nominal P < 0.01, Wilcoxon rank-sum test; multivariate P < 0.05, linear model). Also consistent with the transcriptomic results, we observed by IMC higher expression levels of IFNγ-related proteins in anti–PD-1–secondary resistant SCCHN compared with anti–PD-1–naïve and anti–PD-1–primary resistant SCCHN: higher densities of IDO1⁺CD163⁺ macrophages, CD40⁺CD163⁺ macrophages, and IDO⁺ or VISTA⁺ or B2M⁺ tumor cells (nominal P < 0.05, Wilcoxon rank-sum test; Figs. 4A, 5F and G).

As inhibitory and stimulatory immune checkpoint receptors/ligands may be expressed on immune or tumor cells and contribute to tumor immune escape mechanisms (8), we analyzed some of these markers. As mentioned above, the enrichment score of the exhausted T-cell signature, including PDCD1, CTLA4, LAG3, HAVCR2, and TIGIT genes, was significantly higher in anti–PD-1–secondary resistant compared with anti–PD-1–primary resistant and anti–PD-1–naïve SCCHN (Fig. 5B). We investigated the mRNA expression level of these inhibitory immune receptors individually. Compared with anti–PD-1–primary resistant and anti–PD-1–naïve SCCHN, anti–PD-1–secondary resistant SCCHN had higher mRNA levels of LAG3, HAVCR2 coding for TIM3, PDCD1 coding for PD-1, and TIGIT (Fig. 5H).

To confirm these results, we analyzed the protein expression of LAG3 by IMC. The densities of CD8⁺ T cells, CD4⁺ T cells, and Tregs expressing LAG3 were significantly higher in anti–PD-1–secondary resistant SCCHN compared with primary resistant and naïve SCCHN (nominal P < 0.01, Wilcoxon rank-sum test; Fig. 5I). To exclude that the observed increase in densities of LAG3⁺ cells was due to higher densities of T and other immune cells in secondary resistant SCCHN, we analyzed the proportion of CD8⁺ and CD4⁺ T cells as well as Tregs expressing LAG3. The proportion of CD8⁺LAG3⁺/total CD8⁺ T cells was significantly higher in secondary resistant SCCHN than in primary resistant (nominal P < 0.001, Wilcoxon rank-sum test) and anti–PD-1–naïve SCCHN (P < 0.05; Fig. 5J). Similar results were obtained for the proportions of CD4⁺LAG3⁺/total CD4⁺ T cells and Tregs LAG3⁺/total Tregs (nominal P < 0.05, Wilcoxon rank-sum test; Fig. 5J).

These findings highlight that compared with anti–PD-1–primary resistant or anti–PD-1–naïve SCCHN, anti–PD-1–secondary resistant SCCHN are significantly more infiltrated by T cells and exhibit higher levels of immune activation and exhaustion markers.

Patients with B2M-high tumors have better survival than patients with B2M-low tumors when treated with PD-L1 inhibitors

Finally, as 48 patients in our anti–PD-1–naïve cohort were subsequently treated with PD-L1 inhibitors with or without chemotherapy, we investigated whether B2M genomic alterations or protein expression may predict anti–PD-1 efficacy (Supplementary Table S8). For this subgroup, median OS and PFS were 15.1 months (95% CI, 11–25.1 months) and 4.8 months (95% CI, 3–6.9 months), respectively [median follow-up time: 11.26 months (IQR, 6.57–11.35 months)]. Five anti–PD-1–naïve patients had B2M LOH (3 LOH loss and 2 LOH neutral) and were treated with anti–PD-1. Although limited by the low number of patients, no impact of B2M LOH on OS and PFS could be detected (Supplementary Fig. S5A and S5B). In contrast, OS was significantly better for patients with B2M-high tumors (as described above, tumors with more than 20% of tumor cells positive for B2M expression) compared with patients with B2M-low tumors (HR = 0.29; 95% CI, 0.09–0.91; P = 0.03; Cox model) when treated with anti–PD-L1 therapy (Fig. 6A). PFS was not significantly different between the two groups (HR = 0.52; 95% CI, 0.21–1.30; P = 0.2; Cox model; Fig. 6B). Similar results were obtained using a Cox proportional hazards model integrating anti–PD-L1 regimen type (anti–PD-L1 administered in monotherapy or in combination with chemotherapy), anti–PD-L1 treatment line (first- vs. second-line for R/M disease), HPV status (positive vs. negative), disease extent (locoregional only vs. presence of distant metastasis), performance status (Eastern Cooperative Oncology Group 0 vs. 1 vs. 2), and age (Fig. 6A and B). Response rate was higher in B2M-high patients, but the difference was not statistically significant (P = 0.07, Fisher’s exact test; Fig. 6C). To further assess the potential role of B2M expression level in predicting response to anti–PD-L1–based therapy, we computed the receiver operating characteristic (ROC) curve using a logistic regression model (Fig. 6D). The area under the curve (AUC) was 0.76 (95% CI, 0.549–0.962), confirming the capacity of B2M expression to discern patients likely to respond. In the small subgroup of anti–PD-L1–naïve patients not treated with anti–PD-L1 therapy after tumor biopsy and with survival data available (n = 17), patients with B2M-high tumors seemed to also have an improved OS (HR = 0.32; 95% CI, 0.10–1.02, P = 0.05, Cox model; Supplementary Fig. S5C). This supports the hypothesis that B2M expression may also be prognostic.

Figure 6.

Patients with B2M-high tumors have better survival than patients with B2M-low tumors when treated with PD-L1 inhibitors. A, Kaplan–Meier estimates of OS for B2M-high vs. B2M-low SCCHN treated with anti–PD-1 treatment (left). The P value was from the stratified log-rank test. Stratified Cox proportional hazards model for OS (right). B, Kaplan–Meier estimates of PFS for B2M-high vs. B2M-low SCCHN treated with anti–PD-1 treatment (left). The P value from the stratified log-rank test. Stratified Cox proportional hazards model for OS (right). C, Response rate of anti–PD-L1 treatment. The P value was from Fisher’s exact test. D, ROC curve assessing B2M expression to predict response to anti–PD-1 treatment using logistic regression. AUC: 0.76 (95% CI, 0.549–0.962). E, Kaplan–Meier estimates of OS for CPS ≥1 and B2M-high, CPS ≥1 and B2M-low, CPS <1 and B2M-low or B2M-high (left). The P value was from the stratified log-rank test. Stratified Cox proportional hazards model for OS (right). F, Kaplan–Meier estimates of PFS for CPS ≥1 and B2M-high, CPS ≥1 and B2M-low, CPS <1 and B2M-low or B2M-high (left). The P value was from the stratified log-rank test. Stratified Cox proportional hazards model for PFS (right). ECOG, Eastern Cooperative Oncology Group; ICI, immune checkpoint inhibitor.

As B2M expression was predictive of anti–PD-1 response, we further investigated the relationship between tumoral B2M expression and other predictive biomarkers. To include the maximum number of patients, we analyzed the entire study population. TMB did not correlate with the proportion of B2M⁺ tumor cells and was not different between B2M-low and B2M-high SCCHN (Figure S6A). Clinical CPS data were only available for a minority of patients, so we computed a CPS-like score with the IF data by calculating the proportion of total PD-L1⁺ cells divided by the number of tumor cells. Clinical CPS and the CPS-like score had a high correlation (Spearman’s ⍴ = 0.92, P = 0.001) in the eight patients for whom both scores were performed on the same tumor biopsy (Supplementary Fig. S6B). The CPS-like score showed low correlation with the proportion of B2M⁺ tumor cells (Spearman’s ⍴ = 0.2, P = 0.05), and no difference was observed between B2M-high and B2M-low tumors (Supplementary Fig. S6C). The T cell–inflamed gene expression signature was significantly higher in B2M-high tumors compared with B2M-low tumors (P < 2e-5, Wilcoxon rank-sum test; Supplementary Fig. S6D), consistent with the known regulation of B2M expression by IFNγ (30). The APM was significantly higher in B2M-high tumors (P < 7e−7, Wilcoxon rank-sum test), and we observed a positive correlation with the proportion of B2M⁺ tumor cells (Spearman’s ⍴ = 0.54, P = 4.4e−8; Supplementary Fig. S6E).

Subsequently, we assessed the combination of tumoral B2M expression and CPS-like scores for patient stratification and treatment outcome prediction in the anti–PD-1–naïve patients treated with anti–PD-L1 after biopsy sampling. CPS-like score alone was predictive of PFS (P < 0.05, Cox model) but not of OS (Supplementary Fig. S6F). We compared patients with a CPS-like score ≥1 and B2M-high tumors, patients with a CPS-like score ≥1 and B2M-low tumors, and patients with a CPS-like score <1 and B2M-high or B2M-low tumors: the first group had a better OS and PFS, although this was not statistically significant in our small group of patients (P = 0.08 and 0.07, respectively, Cox model; Fig. 6E and F). TMB and the T cell–inflamed GEP signature were not assessed in combination with B2M expression as none of these patients had a high TMB (>10 mutations/Mb), and T cell–inflamed GEP signature was not predictive of OS and PFS in our cohort (Supplementary Fig. S6G).

Discussion

Understanding resistance to immunotherapy is crucial to better select patients and develop strategies to overcome it. Using a unique cohort of R/M SCCHN who progressed under anti–PD-1–based therapy, we show that anti–PD-1 resistance mechanisms are heterogeneous, involving diverse molecular pathways, and differ between primary and secondary resistance. We found MYC pathway alterations enriched in anti–PD-1–resistant SCCHN compared with anti–PD-1–naïve SCCHN, evidence that T-cell exhaustion may contribute to secondary resistance, and that high B2M expression was associated with improved OS in patients treated with anti–PD-L1–based therapies.

MYC pathway alterations, including amplifications of MYC or MYCL, or MGA mutations, were almost exclusively found in resistant SCCHN. MYC activation has been linked to immunotherapy resistance in preclinical models and other cancer types (32–37). In non–small cell lung cancer (NSCLC), MYC amplifications in pretreatment biopsies correlated with poorer responses to nivolumab or pembrolizumab (32). In patients with urothelial cancer, high MYC expression in pretreatment samples correlated with shorter survival following atezolizumab treatment (33). In SCCHN, B2M expression has also been associated with improved survival for immune checkpoint inhibitor-treated patients (38), and two cases of acquired MYC amplification at tumor progression have been reported after nivolumab (36, 37). Preclinical models showed that MYC activation repressed IFNγ signaling and reduced immune cell infiltration (32–34). Several compounds targeting MYC are being developed and evaluated in preclinical models and early-phase trials. Omomyc, a peptide disrupting MYC–MAX dimerization, was well tolerated and achieved disease stabilization in several patients, whereas MYCi975 demonstrated efficacy in SCCHN preclinical models (39, 40). Although no MYC-targeted drug is approved for R/M SCCHN, these approaches warrant clinical evaluation in this setting. RRAGC is located on the same locus as MYCL and encodes a protein that activates the mTORC1 complex. Alterations in RRAGC have not been previously associated with anti–PD-L1 resistance.

We found that EGFR amplification was enriched in resistant tumors. EGFR alterations, mainly mutations, have already been associated with anti–PD-L1 resistance (41). In glioblastoma, EGFR amplification predicted worse survival under PD-1 blockade (42). This supports the investigation of anti–PD-1 therapy in combination with EGFR inhibitors. In SCCHN, combinations of cetuximab (anti-EGFR), afatinib (pan-HER tyrosine kinase inhibitor), or bispecific EGFR antibodies with PD-L1 inhibitors have been tested and shown promising efficacy (43–45).

We examined the antigen processing and presentation pathway. Apart from more frequent B2M LOH in secondary resistance, alterations in genes implicated in this pathway were not enriched in resistant tumors. B2M LOH did not correlate with B2M protein expression on tumor cells; however, B2M loss and other genomic events in this pathway were associated with reduced B2M expression. Furthermore, although limited by the low number of samples, B2M-high patients had better survival under PD-L1 blockade, consistent with prior findings in R/M SCCHN using digital spatial analysis (38). Further investigations in prospective and larger SCCHN cohorts are required to better study the predictive and prognostic values of B2M expression.

Secondary resistant tumors had higher IFNγ signatures, more stromal and intratumoral CD8⁺ T cells, and increased immune checkpoint expression such as LAG3. These features may reflect either mechanisms of immune escape or prolonged clinical benefit preceding the development of resistance. Similar patterns of chronic IFNγ signaling and T-cell exhaustion have been associated with anti–PD-L1 resistance in different cancer types (8, 46, 47). Transcriptomic analyses of 29 patients with NSCLC with acquired resistance to anti–PD-L1 therapy showed an increase in the expression of IFNγ-related genes at progression, associated with an increase in CD8⁺ T-cell exhaustion signatures (46). These findings support exploring upfront immunotherapy combinations to prevent or delay acquired resistance.

An apparent paradox emerged. Although high B2M expression was associated with survival and immune infiltration, secondary resistant tumors also displayed elevated B2M together with features of T-cell exhaustion and immune checkpoint upregulation. This suggests that B2M may serve as a marker of favorable preexisting immune activation, but under sustained immune pressure, it may also indicate conditions that enable adaptive resistance. Similar dual roles of immune activation markers have been reported in melanoma and NSCLC, in which robust CD8⁺ infiltration promotes initial responses but, through chronic antigen exposure and sustained IFNγ signaling, can ultimately lead to T-cell dysfunction (46).

One of the hypotheses to explain the absence of benefit from the addition of anti–PD-L1 therapy to curative intent chemoradiation is the important radiation dose administered to the regional cervical lymph node (48). We observed that fewer secondary resistant patients had received curative-intent radiotherapy with nodal irradiation, suggesting that the lack of lymph node irradiation may facilitate immune activation and delay resistance. Patients were heterogeneous with respect to line of therapy (first-line vs. post-platinum), treatment regimen (monotherapy vs. chemoimmunotherapy), and tissue of sample collection (e.g., lymph nodes, liver). Although no major differences in the tumor immune microenvironment were detected according to these variables and they were adjusted for in multivariate analyses, such heterogeneity may still have influenced the observed associations.

Our work has several limitations. First, it was a retrospective analysis, albeit conducted on patient cohorts prospectively recruited as part of EORTC clinical trials. Second, unlike many other cancers, R/M SCCHN represents a highly heterogeneous disease. Such heterogeneity that includes clinical and biological features (e.g., smokers vs. nonsmokers, HPV-positive vs. HPV-negative patients, or those with only locoregional recurrence vs. distant metastases) may have influenced our findings. In addition, changes in the standard of care during the inclusion period introduced variability in treatment exposure, with some patients receiving PD-1 inhibitors as first-line therapy in the R/M setting and others after platinum therapy, in line with earlier practice. Finally, our cohort included patients treated with anti–PD-1 either as monotherapy or in combination with chemotherapy, further adding to this heterogeneity. Third, the small sample size of certain subgroups, namely the anti–PD-1–secondary resistant patients, limited the statistical power of some analyses. Lastly, although we had the unique opportunity to analyze biopsies from patients who progressed on anti–PD-1 therapy, a key limitation is the absence of paired pretreatment samples and biopsies from responders. Such comparisons would have enhanced both the statistical power and the biological interpretability of our findings, and in their absence, it remains challenging to determine whether the observed differences truly represent resistance mechanisms or rather treatment-induced changes.

Altogether, our study describes the genomic, transcriptomic, and immune profile of R/M SCCHN at progression on anti–PD-L1 therapy. These findings highlight MYC and EGFR as potential targets, suggest T-cell exhaustion drives secondary resistance, and support B2M as a predictive biomarker of response to anti–PD-1 therapy.

Supplementary Material

Figure S2

Figure S2. Genomic profile of anti-PD1 naïve and anti-PD1 resistant R/M SCCHN, related to figure 1

Figure S3

Figure S3. Some genomic alterations are associated with tumor B2M expression level and high tumor B2M expression is associated with increased T cell infiltration, related to figure 3

Figure S4

Figure S4. Anti-PD1 secondary resistant SCCHN show enhanced immune infiltration and higher expression of immune exhaustion markers, related to figure 4

Figure S5

Figure S5. Patients with high-B2M tumors have a better survival than patients with B2M-low tumors when treated with PD(L)1 inhibitors, related to figure 5

Figure S6

Figure S6. B2M expression is complementary to other biomarkers of survival under PD(L)1 inhibitors, related to figure 5

Supplementary Tables 1-3, 5-8

Supplementary Tables 1-3, 5-8

Supplementary Table 4

Supplementary Table S4: List of oncogenic pathways analysed

Figure S1

Figure S1. Study flow diagram and FFPE vs frozen tumor biopsies comparisons

Data Availability

WES data (FASTQ and VCF files) can be accessed at European Genome-Phenome Archive (EGA) accession numbers: EGAD50000002205 for EORTC-SPECTA and EGAD50000002178 for EORTC-HNCG-1559). RNA-seq data (FASTQ and count files) can be accessed at EGA (accession numbers: EGAD50000002206 for EORTC-SPECTA and EGAD50000002179 for EORTC-HNCG-1559). mIF data (TSV files with proportions of cell types) are available at https://doi.org/10.5281/zenodo.17940577. Code used for analysis and figure generation can be found at https://github.com/ImmucanWP7/immucan_scchn_apd1_resistance.

Authors’ Disclosures

A. van der Elst reports grants from the Belgian National Research Fund (FNRS/Télévie) during the conduct of the study. D. Herrero-Saboya reports that while engaged in the research project, Herrero-Saboya was a PhD student financed by the Institut de Recherches Internationales Servier. B. Bodenmiller reports B. Bodenmiller has founded and is a shareholder and member of the board of Navignostics, a precision oncology spin-off from the University of Zurich. H.S. Hong reports other support from the healthcare organization Merck KGaA, Darmstadt, Germany, during the conduct of the study. P. Bossi reports personal fees and nonfinancial support from Merck, grants and personal fees from Regeneron and BeOne, and personal fees from MSD, LeoPharma, Daiichi Sankyo, GSK, AstraZeneca, Pfizer, and Genmab outside the submitted work. J. Oliveira reports grants from AstraZeneca outside the submitted work. C. Even reports personal fees from MSD, Merck Serono, and Leo Pharma outside the submitted work. S. Lucas reports grants from AbbVie (pharmaceutical company) outside the submitted work. P. Saintigny reports grants from Roche, SMARTCATCH, ADMIR, and Illumina and nonfinancial support from DeepLife outside the submitted work. C. Lefebvre reports personal fees from Servier outside the submitted work. J.-H. Machiels reports other support from advisory board members or speakers with honoraria: Pfizer, Roche, Bayer, Merck Serono, Boehringer, Bristol Myers Squibb, Novartis, Incyte, Cue Biopharma, ALX Oncology, iTEOS, eTheRNA, NEKTAR, F-Star, Seagen, Genmab, Astellas, CureVac, MSD, GSK, Merus, Ipsen, ANAVEON, and Bicara during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

A. van der Elst: Data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing, co-first author. D. Herrero-Saboya: Software, formal analysis, visualization, methodology, writing–original draft, writing–review and editing, co-first author. L. Michon: Software, formal analysis. M. Morfouace: Conceptualization, funding acquisition, methodology, project administration. R. Liechti: Software. P. Devanand: Resources. D. Schulz: Resources. M. Persoons: Resources, data curation. S. Rusakiewicz: Resources. N. Eling: Resources. P. Nicolas: Software. M.-S. Robert: Project administration. S.R. Tissot: Resources. M. Daniel: Resources. B. Bodenmiller: Resources. H.S. Hong: Conceptualization, funding acquisition. R. Galot: Resources. P. Bossi: Resources. J. Oliveira: Resources. F. Estrade: Resources. C. Even: Resources. S. Lucas: Supervision. P. Saintigny: Supervision. L. Martignetti: Supervision. C. Lefebvre: Supervision. J.-P.H. Machiels: Conceptualization, supervision, methodology, writing–original draft, writing–review and editing.