Immunotherapy Response Predictive Score Based on Tumor Microenvironment Profiles for Predicting Immunotherapy Outcomes in Advanced Head and Neck Cancer

Abstract

Objectives

This retrospective study presents an integrative transcriptomic approach for recurrent and/or metastatic head and neck squamous cell carcinoma (R/M HNSCC) by developing an immune response predictive score (IORPS) derived from tumor microenvironment (TME) transcriptomic profiles.

Methods

A total of 30 R/M HNSCC patients treated with pembrolizumab or nivolumab, with available immune TME profiling data, were analyzed. IORPS was constructed based on the cumulative weighting of differentially expressed gene (DEG) expression levels. The predictive performance of conventional biomarkers, individual DEGs, and IORPS was evaluated for immunotherapy response and prognostic outcomes. The clinical relevance of IORPS was further validated using two external cohorts from the GEO database (CLB-IHN: GSE159067 and GHPS: GSE159141).

Results

By comparing immune tumor microenvironment (TME) profiles between good and poor responders, GZMH, IFNG, and FASLG were identified as key DEGs with significantly higher expression in favorable immunotherapy responders. The IORPS, derived from transcriptomic profiling, demonstrated robust predictive accuracy for both immunotherapy response and survival outcomes in patients with R/M HNSCC.

Conclusion

Compared with the variable predictive performance of current biomarkers such as TPS and CPS, IORPS provides improved accuracy and reliability in identifying and stratifying patients most likely to benefit from immune checkpoint blockade therapy.

Introduction

Head and neck squamous cell carcinoma (HNSCC) are the seventh most common malignancy globally, with approximately one million new cases and around 460,000 deaths in 2020.^1,2 In the recurrent and metastatic (R/M) setting, HNSCC patients face a poor prognosis, with overall survival (OS) ranging from 10 to 14 months. ³ While advanced HNSCC generally refers to locally advanced, unresectable disease that may still be treated with curative intent using chemoradiotherapy, recurrent or metastatic HNSCC indicates disease relapse or distant spread beyond curative potential, where systemic therapy becomes the mainstay of treatment. ⁴ Conventional platinum-based chemotherapy, often combined with 5-fluorouracil, taxanes, or cetuximab, provides only limited and short-lived benefit, with significant toxicity that restricts long-term use. ⁵ These limitations underscore the urgent need for more effective and tolerable treatment strategies.^5–7 In Taiwan, the 5-year relative survival rate for stage IV oral cancer patients is 36.17%. ⁸ Together, these data highlight the substantial unmet clinical need in advanced and R/M HNSCC.

Over the past decade, immunotherapy, particularly immune checkpoint inhibitors (ICIs), has transformed the treatment landscape across several cancers, including HNSCC. Two anti-programmed death 1 (PD-1) monoclonal antibodies, pembrolizumab and nivolumab, play critical roles in treating HNSCC. ⁹ PD-1 signaling suppresses T-cell-mediated immune responses, allowing cancer cells to evade the immune system. Anti-PD-1 inhibitors block this suppression, activating the tumor microenvironment (TME) and the immune system. ¹⁰ In HNSCC, the first phase Ib trial, KEYNOTE-012, showed a 38% disease control rate (DCR). ¹¹ Subsequent phase II and III trials, KEYNOTE-055 and KEYNOTE-040, reported a DCR of 35%, with OS of 8 months and a duration of response of around 18 months.^12,13 The CheckMate 141 trial demonstrated a significant survival benefit for nivolumab in R/M HNSCC patients after platinum chemotherapy (hazard ratio [HR] for OS = 0.70; p = .010). ¹⁴ Collectively, these findings established PD-1 inhibitors as a new standard of care for R/M HNSCC.

Following these pivotal studies, attention has shifted toward biomarker-driven approaches to refine patient selection for immunotherapy. Initially, tumor cells expressing programmed death ligand 1 (PD-L1), measured by tumor proportion score (TPS), were evaluated in non-small cell lung cancer. ¹⁵ Tumor cells and immune cells within the TME, assessed by the combined positive score (CPS), became central in guiding immunotherapy decisions.^12,16 In the KEYNOTE-012 study, patients with PD-L1-positive tumors had a significantly higher objective response rate (ORR) compared to PD-L1-negative patients (22% vs 4%; p = .021). ¹⁷ In the KEYNOTE-040 trial, pembrolizumab showed significant benefits in patients with a PD-L1 TPS ≥50%, with a median OS of 11.6 months versus 6.6 months for the standard of care. ¹² However, among patients with a TPS <50%, there was no significant difference in OS (6.5 vs 7.1 months). ⁶ In KEYNOTE-048, CPS replaced TPS as the benchmark for ICI monotherapy or combination therapy. ⁶ Pembrolizumab improved OS versus cetuximab with chemotherapy in patients with CPS ≥20 (median 14.9 vs 10.7 months, HR = 0.60) and CPS ≥1 (12.3 vs 10.3 months, HR = 0.78, p = .0086), and it was non-inferior in the total population (11.6 vs 10.7 months, HR = 0.85). However, even in patients with CPS ≥20 treated with pembrolizumab alone, 77% experienced disease progression within 12 months, suggesting that PD-L1 expression, whether measured by TPS or CPS, may not be an ideal predictor for pembrolizumab effectiveness. Building upon these findings, the phase III KEYNOTE-689 trial demonstrated that perioperative pembrolizumab combined with surgery and standard adjuvant therapy significantly improved event-free survival compared with standard therapy alone in resectable, locally advanced HNSCC, particularly among patients with PD-L1 CPS ≥10. ¹⁸ These results highlight the growing role of immunotherapy in earlier disease settings and reinforce the need for more precise predictive biomarkers beyond PD-L1.

In the CheckMate 141 trial, PD-L1 expression was assessed in tumor cells (TC) and tumor-associated immune cells (TAICs) to evaluate clinical outcomes, including OS and progression-free survival (PFS). A study by Ferris et al in 2016 demonstrated that nivolumab improved OS in patients with tumor PD-L1 TC ≥1% and rare PD-L1+ TAICs, with a median OS of 6.7 months versus 4.9 months for the standard therapy (HR = 0.89). ¹⁴ However, for patients with tumor PD-L1 TC <1%, no significant difference in OS was observed between nivolumab and standard therapy (3.7 vs 4.9 months, HR = 1.09). Median PFS was similar between nivolumab and standard therapy in both groups, indicating limited predictive power of PD-L1 for PFS. ¹⁷ These results suggest that PD-L1 expression alone may not adequately capture the complex immunologic dynamics within the TME, prompting exploration of additional biomarkers.

Tumor mutation burden (TMB), which quantifies the number of mutations per megabase (muts/Mb) in tumor cells, has emerged as another potential predictive biomarker for ICIs. ¹⁹ There is a correlation between TMB and ORR with ICIs across 27 tumor types, including HNSCC. ²⁰ The FDA has approved pembrolizumab for all solid tumors with TMB ≥10 muts/Mb, as measured by the FoundationOne CDx assay. High somatic TMB (top 20% in each histology) was associated with better OS in patients treated with ICIs across cancer types. ²¹ Tumors with mismatch-repair deficiency or high microsatellite instability, such as those with methylation of the MLH1 gene promoter or germline mutations in MLH1 and MSH2, also show high TMB and respond well to ICIs. These tumors are enriched with tumor-infiltrating lymphocytes and PD-L1 expression. ²² However, such genetic alterations are more common in gastrointestinal and endometrial cancers and are rare in HNSCC.^23,24

Although TMB and PD-L1 expression are potential biomarkers, both rely on comprehensive genomic profiling (CGP) or whole exome sequencing (WES), making them costly and limiting their clinical utility in HNSCC. ²⁵ Despite numerous studies, a reliable biomarker for immunotherapy response in HNSCC remains elusive. Therefore, in this study, we developed an integrative transcriptomic approach to identify immune-related biomarkers capable of predicting immunotherapy response and survival outcomes in R/M HNSCC. Using tumor tissue samples from these patients, we aimed to identify improved biomarkers to develop a novel immune response predictive score (IORPS), specifically targeting the immune TME for predicting immunotherapy response and prognostic outcomes. The proposed IORPS was further validated using external GEO datasets, indicating that this novel predictive tool may serve as an alternative prognostic indicator for immunotherapy, potentially improving patient stratification and treatment outcomes in R/M HNSCC.

Materials and Methods

Study Design and Patients

We conducted a retrospective study of Recurrent and/or Metastatic HNSCC (R/M HNSCC) patients treated at single institute. Formalin-fixed paraffin-embedded (FFPE) specimens from 32 R/M HNSCC patients were obtained from the Department of Pathology. The use of human specimens and clinical data for this study was approved by the institutional review board and ethics committee.

This study included R/M HNSCC patients who had received immunotherapy, specifically nivolumab and pembrolizumab. All patients were treated with immunotherapy alone without concurrent chemotherapy or targeted therapy. The inclusion criteria were: diagnosis at age 20 years or older; tumor histology of squamous cell carcinoma (grades 1 to 3); ICD-9 site codes specific to the oral cavity (OC), hypopharynx (HPC), oropharynx (OPC); recurrent and/or metastatic disease; and treatment with immunotherapy. All HNSCC patients underwent surgical resection, and the specimens were collected between 2018 and 2021, ensuring the quality of the FFPE specimens. Baseline demographics, clinical characteristics, treatment timelines, and survival data were extracted from the clinical records using approved protocols. The exclusion criteria included patients who had received prior immune checkpoint inhibitors before recurrence, those with non-squamous histology, incomplete clinical or survival data, inadequate RNA quality for transcriptomic analysis, or concurrent malignancies that could confound immune profiling results. The reporting of this study conforms to the TRIPOD guidelines. ²⁶

Clinical Characteristics and Outcome

Patient medical records were assessed for baseline characteristics and clinical outcomes of study cohort. Clinical characteristics were collected as following: age, sex, risk behavior, primary site of tumor (HPC/OPC/OC), American Joint Committee on Cancer stage at immunotherapy. The treatment response of patients was evaluated by using RECIST 1.1-measurable lesions and classified into four categories: complete response (CR), partial response (PR), progression disease (PD), and stable disease (SD), as we reported in a previous study. ²⁷ During immunotherapy treatment, response status was evaluated every 3 months by computed tomography scan or magnetic resonance imaging image. Patients were categorized into good responder and poor responder based on their response to immunotherapy. Those who achieved PD were classified as poor responders, while patients who achieved CR, PR, SD during the study period were classified as good responders. Further prognostic outcomes, including PFS and OS, were documented every 3 months until the end of the study. Patients were categorized into good responders (CR, PR, or SD) and poor responders (PD) based on their best overall response to immunotherapy, consistent with prior studies using RECIST 1.1 response definitions in head and neck cancer immunotherapy trials.^28,29

Evaluation of Common IO Biomarkers

The evaluation of PD-L1 expression in both tumor and immune cells was conducted collaboratively by two proficient pathologists utilizing the PD-L1 IHC 22C3 pharmDx (Agilent, Santa Clara, CA). The characterization of PD-L1 expression was undertaken through the application of TPS and CPS. The determination of PD-L1 positivity underwent a rigorous assessment utilizing both TPS and CPS methodologies. TPS, assessed via membrane staining of tumor cells, quantified the ratio of PD-L1–expressing tumor cells to the entirety of tumor cells. TPS was scored at predetermined expression levels, including levels of less than 50% and 50% or more. CPS, precisely defined as the count of PD-L1 stained cells (encompassing tumor cells, lymphocytes, and macrophages) as a percentage of the total tumor cell count, multiplied by 100, was scored at predetermined expression levels, including levels of less than 1%, 1% to 19%, and 20% or more. ¹⁶ Furthermore, the evaluation of the additional biomarker, TC, representing tumor PD-L1 membrane expression, was conducted through immunohistochemical testing (Dako North America) utilizing a rabbit antihuman PD-L1 antibody (clone 28-8, Epitomics). ²⁰ Scoring was meticulously performed at predefined expression levels, specifically 1% or more and 10% or more, requiring a minimum evaluation of 100 tumor cells for accuracy.

Evaluation of Immune TME Profiles

To establish an expression signature of the TME, in addition to PD-1 and PD-L1, immune-related genes were assessed using the ACTTME assay, which is performed on a proprietary chip-based multiplex quantitative PCR platform. This assay evaluates immune-related genes and pathways associated with various immune cell types and functions, including active CD4 and CD8 T cells, effector memory T cells, myeloid-derived suppressor cells, regulatory T cells, immune checkpoints, tumor-secreted cytokines, immune response signatures, catalytic activity, activated B cells, natural killer cells, neutrophils, Wnt signaling, hypoxia, stimulatory dendritic cell regulation, and angiogenesis. The test measures the expression of 106 immune-related genes that play critical roles in antigen presentation, immune checkpoint regulation, immune cell population characterization, and other modulatory functions relevant to the tumor microenvironment.

RNA extracted from FFPE samples was reverse-transcribed into complementary DNA. A pre-amplification step was performed to increase the quantity of the target sequences, ensuring adequate templates for quantitative PCR. The prepared mixture of sample and master mix was then loaded onto a QuantStudio 12 K Flex OpenArray chip, which contains 48 subarrays, using the QuantStudio 12 K Flex AccuFill System. The OpenArray chip, once filled, was analyzed using the QuantStudio 12 K Flex Real-Time PCR System.

Transcriptome Profiling

Of the 32 original samples, two were excluded due to failing RNA and data quality control (QC) criteria, leaving 30 samples for analysis. RNA QC required a minimum input of 100 ng, and Data QC necessitated detection of at least two out of three 100-bp integrity control genes (GADPH_100, ACTB_100, B2M_100) and one out of two internal control genes (GUSB or TBP). Samples also needed a scaling factor within the range of −5 to +5, with over 50% of genes detected. For gene expression profiling, Crt values ≥ 35 were capped at 40, and E-values were calculated using Equation (1).

$$E_g=40-Cr{t}_g$$ (1)

where $E_g$ is the E-value of gene g, and $Cr{t}_g$ is the cycle threshold of gene g. $E_g$ represents the relative expression of gene g, where a higher E-value indicates higher gene expression. E-values were normalized to internal control genes to ensure data consistency, and samples with excessive variation were excluded from further analysis. Eventually, 102 genes were detected, with four genes showing no expression across all samples. A heatmap was used to illustrate the E-values of the detected genes across all samples.

Immunotherapy Response Predictive Score (IORPS)

The schematic workflow for IORPS development is summarized in Figure 1. To identify significant differentially expressed genes (DEGs), differential expression analysis (DEA) was performed using gene expression values (E-values), comparing the expression levels between good responders and poor responders to immunotherapy. The detail computation for IOPRS is introduced in the supplementary methods.

Figure 1.

Schematic Workflow for the Development and Validation of Immunotherapy Response Predictive Score (IORPS) Using RNA-seq Data.

The predictive and prognostic performance of DEGs and IORPS was validated using two external cohorts, CLB-IHN (GSE159067) and GHPS (GSE159141), which included patients treated with ICIs.^30,31 Details of the validation cohorts are provided in supplementary methods.

Statistical Analysis

Gene expressions of candidate markers were illustrated using boxplot, and the significance of the microarray gene expression analysis between groups was determined using a Wilcoxon rank sum test. Clinical and pathological features of HNSCC patients were summarized as frequency and percentage, mean and standard deviation, or median and interquartile range (IQR). Receiver Operating Characteristic (ROC) curve analysis was performed to evaluate the predictive accuracy of biomarkers for treatment responses reported as the area under the ROC curve (AUC). PFS and OS were considered as survival endpoint events. The time interval between the date of surgical intervention and the date of these endpoint events was calculated. Survival outcomes were estimated using the Kaplan-Meier method, and log-rank tests were employed to evaluate survival differences between subgroups. Scatter plots with linear fit lines were used to illustrated the relationship between IORPS and observed months to prognostic outcomes after IO. All analyses were two-tailed, p < .05 were considered significant. All analyses were performed using R 4.4.2 software.

Results

Patient Characteristics

A total of 32 R/M HNSCC patients treated with immunotherapy were included in the study; however, only 30 patients passed RNA QC and data QC to provide IO TME profiling data. As a result, the final analysis focused on these 30 HNSCC patients, defined as the KMUH cohort. The exclusion of two samples did not significantly impact the subsequent analysis. The baseline characteristics of all patients and the analyzed subgroup are summarized in supplementary Table S1. The median diagnosis age of the KMUH cohort was 59 years (IQR: 54.0-67.8), and the majority were male (n = 28, 93.3%). Tumors primarily originated from the OC (n = 22, 73.3%), with smaller proportions of patients having tumors in the HPC (n = 4, 13.3%), OPC (n = 3, 10.0%) and HPC + OPC (n = 1, 3.3%). Regarding risk behaviors, 70.0% of the patients (n = 21) had at least one risk factor (alcohol, betel, or cigarette use), with 46.7% (n = 14) exhibiting all three.

In terms of disease stage, most patients were diagnosed at advanced stages: 96.7% (n = 29) were at stages 4A/4B/4C, while 3.3% (n = 1) were at stage 3. Immunotherapy regimens included pembrolizumab (53.3%, n = 16) and nivolumab (50.0%, n = 15), with 10.0% (n = 3) receiving both sequentially. The study concluded on October 7, 2023. Treatment response was assessed for all patients. Among the 30 analyzed patients, 18 (60.0%) were classified as poor responders (immunotherapy duration less than 6 months), and 12 (40.0%) were good responders (immunotherapy duration of 6 months or more). By the end of the study, 24 (80.0%) patients had experienced disease progression, and 21 (70.0%) had died within the study period.

Common IO Biomarkers Exhibit Poor Prognostic Value

Baseline biomarkers such as TPS, CPS, and TC were assessed. Detailed biomarker distributions are also provided in supplementary Table S1. Among the 30 analyzed patients, the median TPS was 20% (IQR: 10% - 60%), the median CPS was 22.5 (IQR: 8.5-68.8), and the median TC was 22.5% (IQR: 10% - 70%). 36.7% (n = 11) of patients showing a TPS≥50, 60.0% (n = 18) of patients showing a CPS≥20, and 80.8% (n = 21) of patients showing a TC ≥ 10. Notably, four patients had missing TC values. The cutoff thresholds were defined according to the criteria established in the KEYNOTE-040, ¹² KEYNOTE-048, ⁶ and CheckMate 141 ¹⁴ studies, respectively. However, the predictive performance of common immunotherapy biomarkers, including TPS, CPS, and TC, using standard clinical cutoff values, yielded unsatisfactory results for predicting favorable IO response in patients with HNSCC (Figure 2). These biomarkers were not ideal for predicting the efficacy of immunotherapy. Figure 2A to C illustrates the poor predictive performance of TPS, CPS, and TC with AUC values of 0.604, 0.595, and 0.640, respectively. In survival analysis, the stratification by these common biomarkers using the typical cutoff values—TPS ≥ 50, CPS ≥ 20, and TC ≥ 10— also did not show any significant difference in PFS (Figure 2D to F) and OS (Figure 2G to H). The analysis strongly indicates that TPS, CPS, and TC have limited ability to predict the immunotherapy efficacy in HNSCC patients, rendering them less reliable biomarkers in this clinical setting.

Figure 2.

Receiver Operating Characteristic (ROC) Plots and Survival Analysis Results for Common Immunotherapy (IO) Biomarkers. (A-C) Receiver Operating Characteristic (ROC) Curves for Tumor Proportion Score (TPS), Combined Positive Score (CPS), and Tumor Cell (TC), Illustrating Their Predictive Ability for Favorable IO Response. (D-F) Kaplan-Meier Plots for Progression-Free Survival (PFS) Based on the Optimal Cutoff Values for TPS, CPS, and TC, Respectively. (G-I) Kaplan-Meier Plots for Overall Survival (OS) According to the Optimal Cutoff Values for TPS, CPS, and TC, Respectively. Note That in Panel (F) and (I), Only 26 Patients are Included due to Four Patients Lacking TC Values.

Immune TME Profiling and Evaluation of DEGs

The gene expression of 102 detected IO TME genes was analyzed and visualized using a heatmap based on E-values (Figure 3A). DEA was performed, and the results are summarized in the volcano plot (Figure 3B). Only three genes, IFNG (fold change [FC] = 4.79, p = .020), GZMH (FC = 3.81, p = .020), and FASLG (FC = 3.26, p = .040), met the significance thresholds (FC > 2, p < .05) and were considered DEGs for subsequent analysis. The predictive performance of these DEGs was evaluated using ROC analysis. As shown in Figure 3D, the AUC values for IFNG, GZMH, and FASLG were 0.745, 0.713, and 0.718, respectively, indicating acceptable predictive performance. DEA results demonstrated significantly higher expression of these DEGs in good responders compared to poor responders, suggesting that higher expression levels are associated with favorable IO responses. Figure 3E illustrates the E-values for IFNG, GZMH, and FASLG between good and poor responders, with IFNG (p < .05) and FASLG (p < .05) reaching statistical significance.

Figure 3.

RNA-seq Analysis Results of the KMUH Cohort Based on the Immune Tumor Microenvironment (TME) Panel. (A) Heatmap of 102 Candidates Immune TME Genes, Annotated and Sorted by IO Response, Comparing Good and Poor Responders. (B) Volcano Plot Displaying the Significantly Differentially Expressed Genes Between Good Responders and Poor Responders. (C) Principal Component Analysis (PCA) Plot Showing the Loadings of IFNG, GZMH, and FASLG among the KMUH Cohort. (D) Receiver Operating Characteristic (ROC) Curve Demonstrating the Predictive Ability of IFNG, GZMH, and FASLG for Favorable IO Response. (E) Boxplot Illustrating the Expression Levels of IFNG, GZMH, and FASLG Between Good and Poor Responders in the KMUH Cohort. Statistical Significance is Indicated as *p < .05, with ns Denoting non-Significance. FC Indicates Fold Change Between Good and Poor Responders. (F-H) Kaplan-Meier Plots for Progression-Free Survival (PFS) Rates Based on the Median Expression Values of IFNG, GZMH, and FASLG, Respectively. (I-K) Kaplan-Meier Plots for Overall Survival (OS) Rates According to the Median Expression Values of IFNG, GZMH, and FASLG, Respectively.

The optimal cutoff values for each DEG were determined using ROC survival analysis. Subsequent survival analyses for PFS and OS were conducted based on these cutoffs. As shown in Figure 3, PFS (Figure 3F to H) and OS (Figure 3I to K) of the high-expression subgroups generally exhibited better survival outcomes compared to the low-expression subgroups. Notably, the high IFNG group (E-value>14.1, p = .023) showed significantly better PFS (Figure 3F), and the high GZMH group (E-value>15.3, p = .023) demonstrated significantly better OS (Figure 3J).

Derivation of IORPS Using DEGs

Since individual DEGs did not provide sufficiently robust outcomes, PCA was applied to capture the combined contribution of each DEG to the overall variance in the KMUH cohort (Figure 3C). PCA revealed that the first principal component (PC1) explained 93.35% of the variance for IFNG, GZMH, and FASLG. The eigenvectors were decomposed from the covariance matrix to retrieve loading scores for IFNG (0.573), GZMH (0.575), and FASLG (0.583) on PC1. These loading scores were subsequently used as weights in the calculation of the IORPS. The computation for IORPS is introduced in the supplementary methods.

Evaluation and Clinical Implications of IORPS

The predictive value and prognostic impact of the IORPS in the KMUH cohort were evaluated, as summarized in Figure 4. Scatter plots with linear fit lines illustrate the relationship between IORPS and the number of months to disease progression (Figure 4A to C) and months since last observation (Figure 4D to F), stratified by IO response, PFS, and OS status. The results indicated that patients in the low IORPS subgroup consistently had shorter observation periods compared to those in the high IORPS subgroup. Additionally, unfavorable outcomes, such as poor IO response, disease progression, and death (Figure 4A to F, red dots and solid lines), were predominantly observed below the threshold line (IORPS < 0.3).

Figure 4.

Immunotherapy (IO) Response Prediction and Prognostic Impact of Immunotherapy Response Predictive Score (IORPS) in the KMUH Cohort. (A-C) Scatter Plots with Linear fit Lines Showing the Relationship Between IORPS and Months to Disease Progression After IO, Stratified by IO Response, Progression-Free Survival (PFS), and Overall Survival (OS) status, Respectively. (D-F) Scatter Plots with Linear fit Lines Depicting the Relationship Between IORPS and Months to Last Observation Since IO Therapy, Stratified by IO Response, PFS, and OS status, Respectively. the Horizontal Dotted Line Represents the IORPS Threshold Value of 0.3. (G) Receiver Operating Characteristic (ROC) Plot Demonstrating the Prognostic Ability of IORPS for Predicting PFS and OS Outcomes. (H-I) Kaplan-Meier Plots for PFS and OS Based on High Versus low Levels of IORPS, Stratified Using a Threshold of 0.3.

Furthermore, the predictive performance of IORPS for PFS and OS was assessed using ROC analysis. IORPS demonstrated an acceptable predictive ability for PFS (AUC = 0.743), but a less satisfactory outcome for OS (AUC = 0.672). Despite this, when the cohort was stratified into high IORPS (≥ 0.3) and low IORPS (< 0.3) groups, significant survival differences were observed. The high IORPS group showed significantly better survival outcomes for both PFS (Figure 4H, p = .032) and OS (Figure 4I, p = .003) compared to the low IORPS group. Overall, these results suggest that IORPS may serve as a surrogate predictor for IO response, PFS, and OS. To address the potential for overfitting, the application of IORPS was further validated using two external cohorts.

External Validation of IORPS on Tumor Response

We validated our findings using two external cohorts retrieved from the GEO database: CLB-IHN (GSE159067, n = 120) and GHPS (GSE159141, n = 51). Both cohorts included data from HNSCC patients treated with immunotherapy targeting PD-1/PD-L1, with tumor responses classified as hot (good responders) and cold (poor responders). Validation results are summarized in Figure 5. First, the TPM values of the identified DEGs were retrieved, and their distribution between hot and cold tumors was illustrated using boxplots (Figure 5A). The results showed consistently higher expression of the DEGs in hot tumors, and all DEGs achieved significant differences between cold and hot tumor responses in both validation cohorts. The computation of IORPS for validation cohorts is introduced in the supplementary methods.

Figure 5.

External Validation Results for Selected Differentially Expressed Genes (DEGs) and Immunotherapy Response Predictive Score (IORPS). (A) Boxplots Illustrating the Expression Levels of IFNG, GZMH, and FASLG in hot Versus Cold Tumor Responses Across Validation Cohorts. the Upper Panel Represents Data from the GSE159067 Cohort (n = 102), and the Lower Panel Represents Data from the GSE159141 Cohort (n = 51). Statistical Significance is Indicated as ***p < .001 and ****p < .0001. (B) Receiver Operating Characteristic (ROC) Curves Demonstrating the Predictive Ability of IFNG, GZMH, FASLG, and IORPS for Favorable IO Response (hot Tumors). the Area Under the Curve (AUC) for Each DEGs and IORPS is Reported. (C) Boxplot Illustrating the Levels of IORPS in hot Versus Cold Tumor Responses in the Validation Cohorts. the Horizontal Dotted Line Represents the IORPS Threshold Value of 0.3.

The predictive performance of each DEG and IORPS was calculated and is summarized in Figure 5B. IORPS demonstrated strong predictive ability, with AUC values of 0.804 and 0.880 in the GSE159067 and GSE159141 cohorts, respectively. Although IORPS had a slightly lower AUC compared to GZMH (AUC range: 0.863-0.946), IORPS provided more consistent results across all cohorts. Notably, GZMH achieved an AUC of 0.713 in the KMUH cohort, which was lower than IFNG (AUC=0.745) and FASLG (AUC=0.718), as shown in Figure 3D. Additionally, the distribution of IORPS in both validation cohorts, visualized using boxplots, showed significantly higher IORPS values in hot tumor responses compared to cold tumor responses (Figure 5C). Overall, IORPS demonstrates promising predictive accuracy for IO response in R/M HNSCC patients treated with immunotherapy.

Discussion

Since 2016, pembrolizumab and nivolumab have emerged as the standard of care for R/M HNSCC. Immunotherapy responses in R/M HNSCC have generally been suboptimal, irrespective of whether administered in the first-line or second-line setting. Nonetheless, a small subset of patients has shown prolonged and effective responses. Given the considerations of cost-effectiveness and precision medicine, the identification of these immunotherapy responders has become increasingly crucial in the context of R/M HNSCC.

Due to the diverse characteristics of each solid tumor, the commonly used biomarkers in the past, including TPS and CPS, also vary in the standards set for different tumors. Head and neck cancer is a type of cancer where treatment efficacy is closely related to the tumor microenvironment, ³² unlike mutation-driven cancers such as lung cancer. ³³ In contrast to cancers where targeted therapies can be identified through genetic testing of cancer cells alone, HNSCC requires a comprehensive assessment of factors including neoantigens, immune cell composition in the microenvironment, and even the surrounding fibrotic stroma to determine the effectiveness of immunotherapy. ³⁴ However, currently available biomarkers, such as PD-L1 expression and TMB, comprehensive genomic profiling (CGP) assays or whole exome sequencing (WES), and the predictive capabilities of these biomarkers are not entirely satisfactory. ²⁵ In our study, we tried to analyzed a novel RNA-based immune profiling assay, ACTTME assay, profiling of 106 genes related to TME in R/M HNSCC patients treated with ICIs.

The ACTTME assay predominantly focuses on assessing the tumor microenvironment, which includes immune-related genes and pathways associated with active CD4 and CD8 T cells, effector memory CD4 and CD8 T cells, myeloid-derived suppressor cells, regulatory T cells, immune checkpoints, tumor-secreted cytokines, immune response signatures, catalytic activity, activated B cells, natural killer cells, neutrophils, Wnt signaling, hypoxia monitoring, stimulatory dendritic cell regulation, and angiogenesis. Surprisingly, the most associated biomarkers and pathways were linked with activated CD8 T cells, nature killer cells, and catalytic activity.

In our study, FASLG, linked to NK cells, exhibited increased expression in individuals who responded positively to ICI. The analysis involved single-cell RNA sequencing and bulk RNA sequencing data from the GEO database, focusing on samples from basal cell carcinoma patients before and after anti-PD-1 therapy. They also found the responders showed higher levels of cytotoxic genes like GZMB and FASLG. ³⁵ Conversely, a separate study investigating the interaction between Fas and Fas ligand (Fas-FasL) in the context of salivary gland cancer indicated that this interaction influences resistance to ICI. ³⁶ Thus, The Fas pathway plays a dual role in modulating immunotherapy within the TME.

A similar scenario arises with IFNG. ³⁷ In our study, IFNG, associated with catalytic activity, demonstrated increased expression in individuals exhibiting positive responses to ICI. Within the TME, a range of immune cells such as activated lymphocytes and NK cells secrete IFN-γ; all nucleated cells respond to IFN-γ due to the expression of the IFN-γ receptor (IFNGR1). ³⁸ However, IFN-γ assumes a dual role in the TME, acting as both an anti-tumorigenic and pro-tumorigenic factor. This duality hinges on the equilibrium between the anti-tumor and pro-tumor signaling of IFN. ³⁸ In a previous study on lung cancer and melanoma cohorts, IFN-γ mRNA emerged as a key predictor for ICI response. The expression of a single IFNG mRNA is indicative of both response and survival outcomes to ICI. ³⁹ The KEYNOTE-012 HNSCC trial investigated a six-gene IFN-γ signature, including CXCL9, CXCL10, STAT1(signal transducer and activator of transcription 1), HLA-DRA(human leukocyte antigen (HLA)-DR-alpha), IFN-γ, and IDO1(indoleamine 2,3-dioxygenase 1) gene expression, in pretreatment biopsies. The aim was to elucidate the intricate relationship between interferons and the response to ICIs. An important link was found between the IFN-γ gene signature and the overall response as well as PFS. ⁴⁰

GZMH has been identified to serve specialized functions within activated cytotoxic T lymphocytes (CTLs), showcasing its integral role in immune responses. Furthermore, recent research suggests that GZMH is commonly integrated into gene signatures associated with CTL activation and effector functions.^41,42 Notably, its presence is frequently observed alongside other granzymes such as GZMA, GZMB, and GZMK, indicating potential collaborative actions in mediating cytotoxic responses of different solid tumors.^43,44 This collective insight underscores the significance of GZMH in orchestrating immune surveillance and cytotoxic activities within the immune system. ⁴² However, there is still a lack of evidence or literature to demonstrate a direct relationship between the expression of GZMH and IO at present.

Previous studies have highlighted the suboptimal suitability of TPS and CPS as biomarkers for identifying patients suitable for immunotherapy. In our cohort, the predictive performance of TPS, CPS, and TC in relation to PFS and OS was deemed insufficient, with limited efficacy for current clinical applications. However, our investigation identified the expression of GZMH, IFNG, and FASLG in the ACTTME assay as promising predictors for patient identification and stratification in R/M HNSCC, enabling a more targeted approach to ICI treatment. Validation using the CLB-IHN and GHPS cohorts confirmed that these biomarkers were significantly higher in patients responsive to immunotherapy, further supporting their predictive potential.

Despite limitations in our cohort, such as a relatively small sample size and short follow-up duration, potential bias arising from subgrouping and the selection of cutoff thresholds for biomarker expression may also influence the predictive outcomes. To minimize this bias, we applied uniform analytical criteria across cohorts and validated the IORPS model externally to ensure robustness. Importantly, IORPS represents an innovative integrative transcriptomic approach that combines immune-related DEGs into a single predictive score, offering a novel perspective distinct from traditional PD-L1 or TMB-based biomarkers. Nevertheless, IORPS shows promise in predicting immunotherapy response with significant clinical implications for R/M HNSCC patients. Its relatively simple RNA-based design makes it both cost-effective and feasible for clinical application, potentially enabling more precise patient selection and early identification of likely responders in real-world settings. By incorporating DEGs from RNA profiling into IORPS, this method provides a more accessible, efficient approach for identifying and stratifying patients for tailored treatments, without the need for time-consuming and expensive NGS. Although our study provides an integrative bioinformatics framework to establish an IORPS based on TME characteristics for R/M HNSCC, the current findings are derived from computational analyses. Further experimental validation, particularly using animal models, is necessary to confirm the biological relevance and mechanistic roles of the identified immune-related signatures.

Conclusions

Compared to single-gene evaluations and common immunotherapy biomarkers such as TPS, CPS, and TC, IORPS offers several advantages for predicting immunotherapy response and survival outcomes in R/M HNSCC patients. While individual DEGs such as IFNG, GZMH, and FASLG demonstrated predictive value, their performance was variable across cohorts, and none consistently outperformed IORPS. Additionally, traditional biomarkers like TPS, CPS, and TC showed limited predictive accuracy for immune response, as evidenced by lower AUC values and lack of significant association with survival outcomes. In contrast, IORPS provided more robust and consistent predictive power across multiple cohorts, with superior stratification of patients into good and poor responders, as well as clearer survival differences in both PFS and OS. Taken together, these findings highlight the innovative nature of IORPS as a comprehensive, RNA-based predictive model that can complement or even surpass existing biomarkers. Its practical applicability and strong predictive performance suggest meaningful clinical utility in optimizing immunotherapy strategies for R/M HNSCC.

Abbreviations

HNSCC

head and neck squamous cell carcinoma

R/M

recurrent and metastatic

OS

overall survival

IORPS

immune response predictive score

TME

tumor microenvironment

DEG

differentially expressed gene

PCA

principal component analysis

ICI

immune checkpoint inhibitors

DCR

disease control rate

PD-L1

programmed death ligand 1

TPS

tumor proportion score

CPS

combined positive score

TC

tumor cells

TMB

tumor mutation burden

TAICs

tumor-associated immune cells

PFS

progression-free survival

ORR

objective response rate

CGP

comprehensive genomic profiling

WES

whole exome sequencing

FFPE

formalin-fixed paraffin-embedded

OC

oral cavity

HPC

hypopharynx

OPC

oropharynx

CR

complete response

PR

partial response

PD

progression disease

SD

stable disease

IQR

interquartile range

ROC

Receiver Operating Characteristic

AUC

area under the ROC curve

FC

fold change