The BIONEXT study is composed of the following two parts ( Figure 1 ).

Part 1: This part included patients with RM HNSCC patients who were treated with nivolumab. Inclusion criteria were age ≧20 years; history of platinum agent administration; pathologically confirmed SCC of the nasal cavity, paranasal sinus, nasopharynx, oropharynx, oral cavity, hypopharynx, or larynx that was recurrent or metastatic and not curable by local therapy; an Eastern Cooperative Oncology Group (ECOG) performance status score of 0 or 1; and at least one tumor lesion measurable per Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1 demonstrated by computed tomography imaging within 28 days of registration. The exclusion criteria were history of ICI therapy or any kind of immunotherapy; and active synchronous or metachronous (within 5 years) cancers except for the carcinoma in situ (CIS) and early esophageal cancer curable by endoscopic resection. Enrolled patients were treated with nivolumab (240 mg every 2 weeks or 480 mg every 4 weeks), and their responses were evaluated every 8 weeks until progressive disease (PD) was detected. Clinical data were collected through the Viedec4 electronic data capture system constructed and maintained by the Clinical Research Support Center (CReS) Kyushu. The endpoint of this study was the identification of a PEX mRNA signature that could segregate non-survivors from long-term (> 2-year) survivors. Pretreatment plasma samples (5 mL), collected from peripheral blood, were preserved at -80°C until assays. Selected pilot samples were subjected to comprehensive RNA-seq analysis for the discovery of candidate PEX mRNA markers, and then the performance of these markers for prognosis prediction was validated by RT-qPCR assays in the entire cohort. All assays were conducted in compliance with the minimal information for studies of extracellular vesicle 2018 protocol (15) in the laboratory of Showa Denko Materials America under strict quality and quantity control anticipating future practical use as a companion diagnostic.

Part 2: This part was designed to confirm that the specific PEX mRNA signature could indeed reflect the TIME of the HNSCC tumors in the identical patient and moreover to elucidate the mechanism of action canceling the effects of nivolumab in non-survivors. Paired tumor and plasma samples were collected from 20 treatment-naïve patients who underwent radical surgery at the National Kyushu Cancer Center. Respective frozen and formalin-fixed paraffin-embedded (FFPE) tumor samples were subjected to mRNA-seq and immunohistochemistry (IHC) to score TIME. Concurrently, the PEX mRNA expression profile of the same patient was evaluated by RT-qPCR in reference to the prognostic biomarker genes established in part 1. Then, patients were stratified into two groups (survivor vs non-survivor signature). Comparing these two cohorts, the TIME score and the biological implication of PEX mRNA signature were investigated.

This study was approved by the Institutional Review Board of the National Kyushu Cancer Center (2019–024), and written informed consent was obtained from all patients before enrolment. This study is registered to the UMIN Clinical Trial Registry: UMIN000037029.

Blood samples, taken within 28 days before nivolumab administration, were immediately centrifuged at 1100xg for 10 minutes and 5 ml of plasma samples were dispensed and snap-frozen at -80°C. Sample collection, preservation, and shipment to Showa Denko America were performed by the SRL Inc. (Tokyo, Japan) under restrict quality and temperature management.

PEXs were quantitatively isolated from plasma using a high throughput ExoComplete isolation tube kit (Showa Denko Materials, Tokyo, Japan), and total RNA was isolated with a MagMax Total Nucleic Acid Isolation Kit (Thermo Fisher, CA) as previously described unless otherwise noted (14). cDNA libraries were prepared using a TruSeq mRNA stranded library kit (Illumina, CA) and sequenced by paired-end read sequencing on a NovaSeq 6000 (Illumina, CA). The obtained raw reads were mapped against the human genome (GRCh38.p13) by hitsat2 and the read counts were obtained by featureCount on a Linux workstation. Differential gene expression analysis was performed by edgeR.

PEX mRNA isolation was conducted as described above. cDNA was synthesized with qScript XLT cDNA SuperMix (Quantabio, MA, USA) following the manufacturer’s protocol. qPCR was performed with SsoAdvanced Universal SYBR Green Supermix (Bio-Rad, CA, USA) in a ViiA 7 Real-Time PCR System (Thermo Fisher Scientific, CA, USA) with the following protocol: 95°C for 10 min, followed by 40 cycles of 95°C for 30 s and 65°C for 1 min and a melting curve analysis. The primer sequences are shown in Supplementary Table S1 . Threshold cycle (Ct) values of the marker candidates were normalized to that of the reference gene (GAPDH) using the delta Ct method.

Human leukocyte antigen E (HLA-E) and programmed death ligand 1 (PD-L1) protein expression levels in the FFPE tumor samples were analyzed using a Ventana Benchmark Ultra slide processor using antibodies against HLA-E (MEM-E/02; Sant Cruz Biotechnology, Inc.) and PD-L1(22C3; PharmDx). The CPS was calculated according to the standard method (5). HLA-E tumor expression was interpreted as strong when more than half of tumor cells was positive, whereas as low when less than half of cells were positive.

RNA extracted from the 17 primary tumor tissues was sequenced on a DNBSEQ-G400 sequencer at Beijing Genomics Institutions (Shenzhen, China). The sequenced reads were aligned to the human reference GRCh38 genome by STAR v2.7.9a with Gencode v38 annotations using the supercomputing system SHIROKANE (University of Tokyo). Transcript-per-million (TPM)-normalized read count tables were generated by RSEM. Downstream analyses were conducted using R v4.1.1. (The R Foundation for Statistical Computing). The IFN-g-signature (the original 6 genes and an expanded 18 genes signature) and the proportions of immune cells in primary tumor tissues were estimated according to the methods in previous reports (6, 7) and CIBERSORTx (https://cibersortx.stanford.edu/) (16). The 17 cases were divided into two groups according to HLA-E expression levels based on the median value. The difference in the IFN-g-signature and the proportions of immune cells between the HLA-E high and low groups were examined by Mann-Whitney U tests. The correlations of the detected marker genes between tissue and PEX-mRNA were examined by Pearson correlation tests.

Data analysis was performed using R version 4.1.1 unless otherwise noted. Statistical significance was determined by a p-value of < 5% derived from ANOVA or Welch’s t test. The performance of the marker candidates was evaluated by the AUC of ROC analysis by R package pROC. The optimum threshold was obtained based on the point of the ROC curve nearest to the top-left corner and used to calculate sensitivity, specificity, positive predictive value (ppv), and negative predictive value (npv) to characterize the performance of marker candidates. Sparse logistic regression was also employed to further validate the predictive values of the biomarkers (17). Survival endpoints used to analyze the candidate biomarkers were visualized using Kaplan-Meier analysis. The log-rank test was applied to test the differences among survival curves. Cox proportional hazards regression models were used to calculate the HR.

Part 1of the study enrolled 111 patients from July 7, 2019, to December 31, 2020, and the clinical data were collected and monitored until July 2022 by CReS Kyushu. Seven patients were excluded due to screening (N = 6) and sampling (N = 1) errors; therefore, the samples and clinical records of 104 patients were utilized for the biomarker assay and survival curve generation. Among them, 7 (6.7%) patients demonstrated a complete response (CR), 12 (12%) had a partial response (PR), 25 (24%) had stable disease (SD), 55 (53%) had progressive disease (PD), and 5 (4.8%) were not evaluated (NE) due to rapid tumor progression. These response rates were similar to those seen in the real world large scale date in Japan (18). The characteristics of the 104 patients are shown in Supplementary Table S2 .

Based on the previous findings that the survival of patients treated by ICI could be stratified by best overall response (BOR) (19), we adopted a standard strategy to initially develop a BOR-predicting biomarker employing receiver operating characteristic (ROC) curve analyses and then to apply this biomarker to prognostic prediction by calculating cumulative survival rates and hazard ratios (HRs) between the marker-selected (i.e., high vs low) cohorts. In preparation for BOR-predicting biomarker exploration, we confirmed the accuracy of BOR for survival prediction in the current cohort (N =104). As shown in Figures 2A, B , the overall survival (OS) rates of patients were well stratified in accordance with BOR; no patients were lost in the CR arm, while extremely poor prognosis was observed in patients with NE, who experienced rapid tumor progression before the first evaluation ( Figure 2A ). Consequently, a substantial difference was found between the curves of responders (N = 19) and non-responders (N = 85) for 2-year OS (93.3% vs. 12.3%, Log-rank test P = 0.00000339; HR: 0.04; 95% confidence interval [CI]: 0.0055-0.293, P = 0.0015079) ( Figure 2B ). However, not only responder (CR+PR), non-responder patients demonstrated long-survival; SD patients revealed a 48.7% of OS at 20 months and PD patients a 20.7% of OS at 2 years ( Figure 2A ), reflecting the fact that a certain portion of patients show durable responses to salvage chemotherapy beyond PD following nivolumab (20). Given that our main goal is to establish an accurate non-survivor-predicting biomarker, these beyond-PD survivors pose a conundrum. This is because, when response-predicting (i.e., responder vs non-responder) biomarkers are applied to survival analyses in this setting, a responder-predicting biomarker with high specificity (score low patients = responders) keeps its power as a survivor-predicting prognostic biomarker, whereas a non-responder predicting biomarker with high sensitivity (score high patients = non-responders) loses its power as a non-survivor-predicting prognostic biomarker, mis-predicting these beyond-PD survivors as non-survivors. Keeping this critical point in mind, we proceeded to the identification of a BOR-predicting PEX mRNA biomarker. We cumulatively collected PEX mRNA sequencing data employing 17 plasma samples of initial phase patients (PR: 6; SD: 5; and PD: 6) when their responses were determined as of November 2020. It is of note that these 6 PR patients were > 2-year survivors (i.e., good responders). Then, we selected candidate BOR-predicting PEX mRNA, adopting a less restricted marker-selecting condition not confining the comparisons of groups between responders and non-responders, thus if they met one of the following criteria: 1) genes that were differentially expressed among the BOR categories (PR vs PD, PR vs SD/PD, and PR/SD vs PD) (P < 0.05), 2) genes with |log(fold change)| > 1.5, 3) genes with high area under the curve (AUC) values (> 0.7) in the ROC analyses for detection of PR vs PD, PR/SD vs PD (AUC1) and PR vs SD/PD (AUC2), or 4) genes identified as potential biomarkers in previous ICI studies (6, 8, 21, 22) or with high |log(fold change)| values in the present study. With these less broad criteria, the top 20 genes, TAF4B, TESK2, MFSD8, RABL2B, ZNF480, FAM76A, TGIF1, TNFRSF13C, CTSW, LOC283788, SLC25A13, HLA-DQA1, COL10A1, MPIG6B, RPL23AP7, MSH2, CD3D, TCF7, HLA-E, and HLA-DRA were selected as candidate BOR-predicting biomarkers for further analyses ( Figure 2C ; Supplementary Table S3 ).

Employing these candidate BOR-predicting PEX mRNAs, their powers for response prediction (i.e., responder vs non-responder) were investigated by RT-qPCR assays in the entire cohort (N = 104). To normalize the PEX mRNA data, two representative reference genes, ACTB and GAPDH, were added to assays. Interestingly, they demonstrated significantly (ACTB, P < 0.001; GAPDH, P < 0.05) higher expression (i.e., raw threshold cycle value) in the non-responder than in the responder. Assuming that these increases may reflect the vigorous total exosome production from aggressive cancer cells as confirmed in previous studies (10), we adopted ACTB, which had a greater difference, as one of the candidate biomarker PEX mRNAs, and used GAPDH as the reference gene. We then compared the expression levels of GAPDH-normalized 21 PEX mRNAs between responders and non-responders and their response-predicting powers were measured by the values of AUC in the ROC curve analyses and their optima thresholds were determined by the point nearest to the top-left corner on the ROC curve. The top 6 genes with the high AUCs, HLA-E, ACTB, MPIG6B, RABL2B, TNFRSF13C, and ZNF480, were selected as putative response-predicting biomarkers ( Supplementary Table S4 ). PEX mRNAs that were increased in the non-responders (HLA-E, ACTB, MPIG6B and TNFRSF13) were considered as non-responder-predicting markers, while those that were increased in the responders (RABL2B, and ZNF480) were considered as responder-predicting markers ( Figure 3A ). Of note, the GAPDH-normalized ACTB PEX mRNA remained significantly higher in non-responders, supporting our hypothesis. The AUC of these PEX mRNAs ranged from 0.593 to 0.729. For comparison, we calculated the AUC of the neutrophil-to-lymphocyte ratio (NLR), a proposed non-responder-predicting biomarker of ICI (23, 24), and found it was 0.591 ( Supplementary Table S4 ; Figure 3A ). The performance of individual PEX mRNAs was better than that of the NLR, but the values were not sufficiently high for clinical use. We then employed a simple algorithm to develop a better signature for response prediction by the combination of multiple PEX mRNA markers and the NLR. With the intent to generate non-responder-predicting combinations, we assigned 1 point if the expression of a non-responder-predicting gene or NLR exceeded the threshold or a responder-predicting gene fall below the threshold using the best threshold value (i.e., the point of the highest sensitivity and specificity) of each marker determined by the ROC curve analysis ( Supplementary Table S4 ), and the points were averaged for various marker combinations. The score ranged between 0 and 1, and a score of 0 indicated that no marker in the combination predicted a non-responder, while a score of 1 indicated that all markers in the combination predicted a non-responder. To obtain the best combination of markers, we tested all the possible combinations of the top 6 markers and the NLR and identified the top 10 combinations with higher AUCs (ranging from 0.793 to 0.812) ( Table 1 ; Figure 3B ). In the comparison of responders and non-responders, the scores of these combinations demonstrated more significant differences (P < 0.0005) ( Figure 3B ) than the mean expression of individual 6 PEX mRNAs and NLR, in which only HLA-E, ACTB (P < 0.05) and NLR (P < 0.005) demonstrated significant differences ( Figure 3A ). Notably, all the top 10 combinations included HLA-E, which may suggest its importance for response prediction ( Table 1 ).

In our final assay, we investigated whether these non-responder-predicting combinations can serve as prognostic biomarkers for the prediction of non-survivors. Kaplan-Meier curves of patients were generated according to the thresholds of combinations 1-10 ( Table 1 ). In combination 1, 2, 4, 7, 9, and 10, patients with high non-responder scores (above the threshold) demonstrated significantly (P = 0.0002348-0.0238) higher HRs (2.09-2.878) ( Figure 4A ). Strikingly, in the most promising (i.e., high HR) combinations (9 and 10), the OS of the patients with high non-responder scores demonstrated a sharp drop towards 0% at 2 years ( Figure 4B ), while that of patients with low non-responder scores demonstrated an approximately 60% 2-year OS and a tendency to plateau after 20 months. Considering the highest HR and the lowest P value, we determined to adopt the combination 9 as a prognostic biomarker of nivolumab.

In part 1 of our study, we identified a prognostic biomarker combination (HLA-E, ACTB, MPIG6B, RABL2B, TNFRSF13C, ZNF480 and NLR) that could precisely predict non-survivors treated with nivolumab. We then proceeded to part 2 of the study to confirm that the combination of 6 PEX mRNAs and NLR indeed reflect the TIME and, more importantly, to find a biological clue for the development of novel strategies for non-survivors ( Figure 1 ). For part 2, 20 paired blood, plasma and tumor samples were collected from the treatment-naïve HNSCC patients who underwent radical surgery at National Kyusyu Cancer Center. This is mainly because it is often difficult to obtain appropriate tumor samples from the patients with R/M HNSCC. Blood samples were used for the measurement of NLR. Plasma samples were subjected to PEX mRNA assay and tumor samples were to RNA-seq and IHC. Sufficient tissue amounts for RNA-seq were not obtained for 3 frozen tumor samples; thus, 17 tumor samples were subjected to the mRNA analyses, 20 tumor samples were subjected to the IHC analysis, and 20 plasma samples for PEX mRNA assay. We first measured the expression levels of GADPH-normalized 6 PEX mRNAs by RT-qPCR and the levels of mRNAs in the corresponding tumors by RNA-seq to examine their correlations. Consistent with the previous finding that only specific genes demonstrated significant correlations (25), PEX HLA-E mRNA showed a near-significant (P = 0.052) correlation with tumor HLA-E mRNA among the 6 genes ( Figure 5A ). In view of this positive tendency, we compared the expression levels of PEX HLA-E mRNA and HLA-E protein in the tumors and found a significant association (P = 0.0191) ( Figure 5C ). Collectively, the high PEX HLA-E mRNA expression appears to reflect the high HLA-E mRNA transcription and protein translation in the corresponding tumor. We then attempted to stratify the 20 patients into score high candidate non-survivors and score low candidate survivors based on the biomarker combination 9 established in part 1 of the study ( Table 1 ; Figure 1A ). However, interestingly, all 20 patients were grouped with survivor signature, because in the blood and plasma samples obtained from the treatment-naïve patients the NLR and the mean PEX mRNA expression levels of 6 PEX mRNAs except for TNFRSF13C indicated favorable response patterns compared to the RM samples ( Figure 5B ); HLA-E, ACTB, and MPIG6B (non-responder genes) were lower and RABL2B and ZNF480 (responder genes) were higher. This result is consistent with the fact that the TIME of treatment-naïve tumor is more tumor-eliminating compared to the exhausted TIME of RM tumor, warranting the efficacy of this biomarker as a monitor of the TIME. Given the prominent role of HLA-E repeatedly identified in the present study and its importance as a target of immunotherapy (i.e., therapies targeting the HLA-E-NKG2A immune checkpoint) (26, 27), we alternatively utilized the mean value of PEX HLA-E mRNA to stratify the 20 patients. We compared the status of immune parameters (PD-L1 CPS score, IFN-γ-related signature score, and CIBERSORT-derived infiltrating immune cell levels) (6, 7, 16) between PEX HLA-E mRNA high (N = 10) and low (N = 10) patients. The CPS (P = 0.6242) and IFN-γ-related signature (P = 0.1802) did not show significant correlations with the levels of PEX HLA-E mRNA. However, the number of activated natural killer (NK) cells determined by CIBERSORT were significantly (P = 0.024) abundant in the tumors of patients with high PEX HLA-E mRNA ( Figure 5D ). It is known that HNSCC is the most immune-infiltrating cancer types across the solid tumors (28) and these tumor-infiltrating NK cells and CD8⁺ cytotoxic T lymphocytes (CTL) strongly express NKG2A and PD-1 (27). Considering the positive correlation of PEX HLA-E mRNA and HLA-E protein expression confirmed above, the effects of PD-1blockade by nivolumab may be canceled by HLA-E-NKG2A check point in patients with high PEX HLA-E mRNA, as illustrated in Figure 5E . Thus, the combination of clinically usable anti-NKG2A antibody (i.e., monalizumab) and nivolumab may be useful for the candidate non-survivors predicted by the pre-treatment biomarker combination indicating high PEX HLA-E mRNA.