1 Introduction
Nasopharyngeal carcinoma (NPC) is a unique type of head and neck cancer originating from the mucosal epithelium within the nasopharynx, most commonly occurring in the roof and lateral walls of the nasopharynx, particularly in the pharyngeal recess [
1]. According to the World Health Organization, NPC can be categorized into three pathological subtypes, including keratinizing, non-keratinizing, and basaloid squamous cell carcinoma. Non-keratinizing NPC can be further divided into differentiated and undifferentiated carcinoma [
2]. Although globally considered as an uncommon cancer, NPC is frequent in East and South-East Asia, especially in southern China where the incidence rate is more than 20 times higher than the global average, showing extremely unbalanced distribution [
3]. The 2022 GLOBOCAN data showed that there were more than 120 000 new cases of NPC globally [
4], among which 51 000 new cases occurred in China with 28 400 deaths [
5]. In non-endemic areas, the incidence of NPC follows a bimodal distribution with an initial peak in adolescents and young adults (ages 15–24 years), and a second peak in individuals over 65 years old. In endemic areas where non-keratinizing NPC predominates, the incidence starts to rise after the age of 30, peaks between 40 and 59 years, and then gradually declines [
6]. The male-to-female incidence ratio is 2.63:1 [
7], and females generally have a better prognosis than males [
8].
NPC has a complex and unique etiology involving viral infection, particularly Epstein-Barr virus (EBV). EBV is a key factor in NPC tumorigenesis, with a prevalence of 100% in the non-keratinizing subtype that accounts for over 95% of NPC in endemic areas. The virus infects epithelial and B cells, integrating into host DNA, preventing apoptosis and promoting cell growth. Environmental risk factors, such as elevated nickel levels and dietary patterns (e.g., salted fish consumption), also contribute to the disease. Additionally, NPC shows familial and racial clustering, suggesting a genetic predisposition [
1]. Exposure to environmental carcinogens induces genetic alterations, including deletions on chromosomes 3p and 9p, which deactivate tumor suppressor genes and increase the susceptibility of nasopharyngeal cells to EBV infection. EBV-transformed nasopharyngeal epithelial cells undergo genomic instability, DNA methylation, and activation of cancer-related pathways, fuelling NPC progression. Mutations in NF-κB and PI3K-MAPK pathways, along with EBV gene expression, contribute to immune evasion and metastasis [
9,
10].
Approximately 70% of patients with newly diagnosed NPCs present with advanced disease [
11]. For these patients, standard treatment modalities, including radiotherapy and chemotherapy, have significantly improved outcomes. However, 20%–30% of patients could still experience recurrence or metastasis [
12], resulting in a poor prognosis, which calls for novel treatment strategies. Once called lymphoepithelioma due to the extensive infiltration of lymphocytes, NPC is characterized by high programmed death ligand 1 (PD-L1) expression, seen in 83%–92% of patients in either tumor cells or tumor-infiltrating immune cells [
13,
14]. These features make it a promising candidate for the therapeutic use of antibodies targeting the programmed cell death protein 1 (PD-1) molecule. However, though the introduction of immunotherapy has achieved substantial benefits in the treatment of NPC, a subset of patients still does not derive benefit from immunotherapy. Therefore, predictive biomarkers are imperative for the selection of appropriate patients and maximizing therapeutic efficacy. While PD-L1 expression has been the most widely studied biomarker, it alone does not fully predict responses in NPC patients. Additional biomarkers, such as EBV DNA, tumor mutational burden (TMB), components in the tumor microenvironment (TME), and factors that may influence host systemic immunity such as metabolism and gut microbiota, are gaining recognition. A critical challenge lies in distinguishing predictive biomarkers (which indicate likelihood of response to a specific therapy) from prognostic biomarkers (which indicate overall disease outcome, independent of therapy). To elucidate the predictive value of a biomarker, rigorous randomized controlled trials (RCTs) with appropriate control groups are indispensable.
In this review, we summarized recent advances in immunotherapy for NPC and studies on biomarkers that were associated with treatment response or long-term survival benefit from immunotherapy. Some of these biomarkers have been validated as predictive of immunotherapy benefit in RCT cohorts, while others require further validation of their predictive value. We also summarized the challenges and future directions of biomarker studies, hopefully facilitating the development of predictive biomarkers for immunotherapy that can eventually enter clinical practice.
2 Immunotherapy for nasopharyngeal carcinoma
The treatment of NPC has evolved significantly over time, driven by advances in radiotherapy and the growing understanding of its biological behavior (Fig. 1). For non-metastatic NPC, radiotherapy has always been the cornerstone of treatment, given the radiosensitive nature of tumor and the challenging anatomical location of the nasopharynx that limits surgical options. Initially, 2D radiotherapy was used, but it was eventually superseded by 3D conformal radiotherapy and, more recently, intensity-modulated radiotherapy (IMRT), which has improved locoregional control and survival rates while reducing side effects [
15]. Further innovations in radiation techniques, such as IMPT and carbon ion therapy (IMCT), offer even more precise dosimetry, minimizing radiation exposure to normal tissues [
16,
17]. For early-stage NPC (stages I and II), radiotherapy alone is often sufficient, though chemotherapy may be added for high-risk features. The 10-year overall survival rates ranged from 87.1%–100% in those patients [
18]. In locoregionally-advanced NPC (LA-NPC, stages III and IVA), a combined-modality approach involving concurrent chemoradiation (CCRT) has become the standard of care, significantly improving survival and reducing the risk of metastasis [
19]. Studies have confirmed that induction chemotherapy (IC) can demonstrate significant survival benefit rather than adjuvant chemotherapy, and gemcitabine plus cisplatin (GP) is one of the recommended regimens [
20–
25]. Metronomic capecitabine also shows promise as an adjuvant treatment for LA-NPC [
26]. For recurrent or metastatic NPC (RM-NPC), chemotherapy is the mainstay of treatment. In recent years, immune checkpoint inhibitors (ICIs) have emerged as a promising addition to NPC treatment, significantly improving long-term survival when combined with chemotherapy in the treatment for RM-NPC, and are now being investigated for the treatment of LA-NPC.
2.1 Mechanisms under anti-PD-(L)1 immunotherapy
T cell activation requires three key signals. The first signal originates from antigen recognition, where the T cell receptor (TCR) engages with antigenic peptides presented by major histocompatibility complexes (MHC) on antigen-presenting cells (APCs). The second signal involves co-stimulatory or co-inhibitory pathways that modulate T cell responses. The third signal is provided by cytokines in the extracellular microenvironment [
27]. Among co-inhibitory regulators, immune checkpoint molecules such as PD-L1, cytotoxic T lymphocyte-associated antigen 4 (CTLA-4), T cell immunoglobulin mucin 3 (TIM-3), and lymphocyte-activation gene 3 (LAG-3) play critical roles in suppressing immune activity [
28]. For example, PD-L1 expressed on tumor cells or APCs binds to PD-1 on T cells, triggering a cascade of immunosuppressive effects—including T cell apoptosis, diminished cytokine release, impaired cytotoxic function, and antigen tolerance—thereby enabling tumors to evade immune surveillance. ICIs counteract this evasion by blocking PD-1/PD-L1 interactions, restoring T cell-mediated tumor targeting and destruction [
29]. The clinical impact of ICIs has been transformative. Following the landmark approval of ipilimumab (anti-CTLA-4) for metastatic melanoma over a decade ago, at least eight ICIs targeting PD-1/PD-L1 or other checkpoints, along with 35 combination therapies incorporating ICIs, have been approved across multiple cancer types [
30]. Notably, anti-PD-(L)1 agents have become conventional therapeutic approaches in the treatment of solid tumors [
31], demonstrating durable responses and reshaping treatment paradigms in oncology.
2.2 Immunotherapy for RM-NPC
Immunotherapy has revolutionized the treatment for RM-NPC, particularly with anti-PD-1/PD-L1 agents. Three recent randomized phase 3 trials—JUPITER-02 [
14,
32], CAPTAIN-1st [
33], and RATIONALE-309 [
34]—have reported similar significant improvements in progression-free survival (PFS) with the combination of anti-PD-1 therapy with GP chemotherapy for the first-line treatment of RM-NPC (Table 1). These findings supported the approval of toripalimab in combination with GP for the first-line treatment of RM-NPC by the United States Food and Drug Administration (FDA) in 2023. In addition, shortly after the National Medical Products Administration (NMPA) of China granted the marketing authorization to tagitanlimab, an anti-PD-L1 monoclonal antibody, for the second-line and beyond treatment of RM-NPC based on a single-arm, phase 2 study [
35], NMPA approved tagitanlimab in combination with GP for the first-line treatment of RM-NPC based on a randomized, phase 3 trial (NCT05294172) in 2025. This is the world’s first anti-PD-L1 antibody approved for the treatment of NPC.
In the setting of second-line and subsequent treatment for RM-NPC, POLARIS-02, a single-arm phase 2 study, displayed durable anti-tumor activity of toripalimab, with an objective response rate (ORR) of 20.5%, a disease control rate (DCR) of 40%, and a median overall survival (OS) of 17.4 months [
36]. Based on this study, the FDA approved toripalimab for the subsequent-line treatment of RM-NPC in 2023. A meta-analysis including 12 studies of PD-1 inhibitors demonstrated an ORR of 23% [
37]. However, KEYNOTE-122, a randomized phase 3 study evaluating pembrolizumab versus chemotherapy in patients with previously treated RM-NPC, did not observe significant OS improvement [
38], suggesting that anti-PD-1 monotherapy might have limited efficacy in second-line treatment for RM-NPC. A phase 2 study examining the efficacy of nivolumab plus gemcitabine demonstrated improved PFS and OS [
39]. Other combination regimens have been under investigation, such as the combination with ICIs of different targets [
40,
41], anti-angiogenesis agents [
42–
45], antibody-drug conjugates [
46], epigenetic drugs [
47], and EBV-specific T cell immunotherapy [
48]. Whether immunotherapy combined with other therapy can truly enhance survival benefits for RM-NPC patients who have failed prior treatment remains to be determined through further study.
2.3 Immunotherapy for LA-NPC
Based on the impactful results in RM-NPC, ICIs have been further tested for LA-NPC treatment, with promising results from several single-arm phase 2 trials showing the safety and favorable outcome of PD-(L)1 blockade combined with chemoradiotherapy [
49–
52]. The CONTINUUM study, which compared sintilimab combined with IC and CCRT followed by sintilimab adjuvant therapy versus standard treatment (IC + CCRT), was the first phase 3 randomized trial showing that the addition of anti-PD-1 to standard chemoradiotherapy significantly improved event-free survival (EFS) in the definitive treatment of patients with LA-NPC [
53]. Results from other randomized controlled trials of PD-1 inhibitors also demonstrated a significant survival benefit when added in a so-called sandwich approach (induction + adjuvant phase) or in the adjuvant phase (Table 2) [
54–
56]. Future studies are needed to determine the optimal way of the combination of anti-PD-1 with chemoradiation. Several ongoing trials aim to evaluate the benefit of other kinds of immunotherapy for LA-NPC, including antibodies targeting other immune checkpoints, bispecific antibodies, ICIs combined with anti-angiogenesis therapy, and immune cell therapy [
57]. Some of them have also revealed promising preliminary results (NCT04447326 [
58]), but the final results of these trials are still pending. Nevertheless, with all the promising results mentioned above, it is believed that immunotherapy will have the potential to reshape the therapeutic landscape for NPC, and better predictive tools to select patients who can benefit from immunotherapy will be imperative.
3 Predictive biomarkers widely studied in other tumors
Several predictive biomarkers have been proposed for anti-PD-1 immunotherapy, among which PD-L1 expression, tumor mutational burden (TMB), and microsatellite instability-high or mismatch repair deficiency (MSI-H/dMMR) have been validated in many types of solid tumors and have been approved by the FDA. However, the predictive roles of these markers have not been well confirmed in NPC.
3.1 PD-L1 expression
Given that the PD-1 pathway is likely a key mechanism of immune evasion, PD-L1 expression on tumor cells or immune cells emerged as a potential candidate biomarker for predicting the efficacy of immunotherapy. Currently, various methods have been proposed for assessing PD-L1 expression levels. KEYNOTE-001 was a pioneering clinical trial exploring PD-L1 expression as a potential biomarker for evaluating the efficacy of immunotherapy in non-small cell lung cancer (NSCLC), in which PD-L1 expression was detected using immunohistochemistry (IHC), and evaluated using the tumor proportion score (TPS) defined as the percentage of tumor cells (TC) with membranous PD-L1 staining [
59]. Since immune cells (IC) also express PD-L1, the combined positive score (CPS) taking account of PD-L1 expression on both tumor and immune cells was proposed and deemed a promising predictive biomarker for immunotherapy efficacy in head and neck squamous carcinoma (HNSCC) [
60].
In regard to NPC, however, the predictive value of PD-L1 expression is not well established. First, the detection methods and reported positivity rates of PD-L1 expression vary considerably across different studies, showing relatively poor consistency among these methods. Studies have shown that using different antibodies (e.g., 22C3, 28-8, SP263, SP142) and testing platforms for PD-L1 detection may lead to significant discrepancies. For instance, 22C3, 28-8, and SP263 exhibit relatively high concordance in detecting PD-L1 expression in tumor cells, whereas SP142 demonstrates lower consistency compared to other antibodies [
61]. In addition, PD-L1 detection results are influenced by factors such as evaluation criteria, platform-specific variables, tumor heterogeneity, and relatively high discordance among pathologists for scoring PD-L1 expression in immune cells as against in tumor cells [
61,
62]. These factors may explain the fact that subgroup analyses from large-scale RCTs in the setting of NPC have showed inconsistent and sometimes opposite results. In JUPITER-02 and RATIONALE-309 studies, immunotherapy showed consistent improvement of PFS regardless of PD-L1 expression status, indicating that PD-L1 expression might be just a prognostic factor rather than an effective predictive biomarker. In KEYNOTE-122 study, patients with CPS < 10 benefited from ICI, while in CONTINUUM study, the interaction analysis assessing treatment effect on EFS in different CPS subgroups turned out to be insignificant (Table 3). Admittedly, the settings of these trials are not exactly the same, but it still emphasizes the urgent need of a standardized PD-L1 detection method to ensure the accuracy and comparability of results, thereby improving the study of its utility in guiding clinical treatment decisions.
3.2 Tumor mutational burden
Tumor mutational burden (TMB), defined as the total number of mutations present in a tumor specimen, is another broadly investigated predictive marker that has been reported to be reflective of tumor immunogenicity and significantly correlated with anti-PD-1 therapy response in various cancers, such as lung cancer and melanoma [
63]. TMB can be calculated using whole-exome sequencing (WES) or next-generation sequencing (NGS) with comprehensive gene panels. Besides, recent advances in blood-based estimation of TMB using circulating tumor DNA (ctDNA) also show consistent and promising results [
63].
However, the implementation of TMB as a predictive biomarker in NPC is limited. First, NPC is not regarded as a mutation-driven cancer since few hotspot mutations were found [
64], with a reported median TMB ranging from 0.95 to 3.05 mutations/megabase (Mb) [
36,
65], far below the cut-off of 10 mutations/Mb specified in the FDA approval. Indeed, despite that some single-arm trials proposed TMB to be predictive, others showed no correlation of TMB with ICI response or survival in NPC [
36,
39,
40,
42,
66]. Second, there is no consensus regarding the optimal cut-off of TMB to be used as a predictive biomarker. Most studies use the median TMB as data cut-off, and the reference values vary greatly across different studies, making it difficult to establish a standardized methodology. Wang
et al. demonstrated the nonlinear relationship between TMB and ICI response, and proposed a mixed-endpoint analysis model to determine the optimal TMB subgrouping. However, the model-derived TMB categorization was still not associated with immunotherapy in the NPC cohort [
67].
Given the unsatisfactory clinical use of TMB in some types of cancer, various modified TMB-related biomarkers have been developed [
68]. For instance, identifying neoantigen clones with the ability to induce immune elimination also shows potential to better predict the efficacy of immunotherapy. Su
et al. developed an immunoediting-based optimized neoantigen load (ioTNL) model that could predict the response and prognosis of immunotherapy in several types of cancer including NPC [
69]. However, similar to TMB, the best detection panel and the optimal cut-off value need further corroboration.
3.3 Microsatellite instability-high
Microsatellite instability (MSI) in tumor DNA is defined as the presence of alternate-sized repetitive DNA sequences that are not seen in the corresponding germline DNA, and is a marker of genomic instability. It is caused by a deficient DNA mismatch repair system, involving the loss of function of mismatch repair (MMR) proteins including MLH1, MSH2, MSH6, and PMS2. IHC is utilized to detect the expression levels of MMR proteins and classify tumors as deficient or proficient in the MMR system (pMMR) accordingly. MSI is assessed using polymerase chain reaction (PCR) to detect a panel of five markers and comparing marker length between normal and tumor tissue. Tumors with instability at two or more of these markers were defined as being of high microsatellite instability (MSI-H), whereas those with instability at one repeat or showing no instability were defined as low MSI (MSI-L) or microsatellite stable (MSS) tumors, respectively [
70]. MSI-H generally occurs as a subset of high TMB. The vast majority (97%) of MSI-H samples also have high TMB (defined as ≥ 10 mutations/Mb). However, the converse was not true; only 16% of samples with high TMB are classified as MSI-H [
71]. MSI-H is a broadly recognized predictive biomarker of ICI response that performs well across several cancer histologies, especially in colorectal carcinoma [
70]. However, as is the case with TMB, the prevalence of dMMR or MSI-H is quite infrequent (~2%) in NPC [
72,
73], hindering its use as a practicable predictor of ICI benefit.
4 Biomarkers in the tumor microenvironment
Given that immunotherapy efficacy relies on complex interactions among cytotoxic cells, regulatory cells, and tumor cells, tumor specimen-based studies have identified immune cells within the TME and tumor-intrinsic features influencing immune evasion or immune-mediated killing as potential predictive markers for immunotherapy (Fig. 2).
4.1 Immune cell subpopulations
Many biomarker studies have been conducted taking advantage of the published clinical trial cohorts of advanced NPC. In the biomarker analysis from a phase 2, single-arm trial that evaluated the anti-tumor activity of camrelizumab, an anti-PD-1 antibody, in pretreated RM-NPC, Yang
et al. found that patients with durable clinical benefit (DCB) had higher density of MHC-II
+ cell in stroma than patients without DCB, resulting in better patient selection to receive immunotherapy [
74]. In a phase 3, randomized trial testing tislelizumab or placebo plus chemotherapy in treatment-naïve RM-NPC, immunologically “hot” tumors with an activated dendritic cell (DC) signature were found to be associated with PFS benefit from tislelizumab-combined chemotherapy, which may help identify patients who might benefit most from immunochemotherapy treatment [
34]. These findings suggest that antigen presentation machinery may play an important role in eliciting anti-tumor immunity and foster tumor elimination during anti-PD-1 treatment in patients with RM-NPC. In the setting of LA-NPC, biomarker analysis of the recently published phase 3, randomized controlled CONTINUUM study demonstrated that the frequency of proliferating regulatory T cells (Ki67
+ Tregs) in total T cells was predictive of anti-PD-1 immunotherapy benefit [
75], suggesting that immunosuppressive factors in the TME could also be predictive and might be targeted to sensitize immunotherapy.
4.2 Spatial structures
Immune cells can form complex spatial structures in the TME where they interact with each other to regulate the anti-tumor immunity. For example, the relationship between tertiary lymphoid structure (TLS) and B cell responses and their roles in predicting ICI response has already been validated in other cancer types like melanoma, renal cell carcinoma, and sarcoma [
76–
78]. In a phase 2 trial to evaluate the safety and activity of camrelizumab plus apatinib in platinum-resistant NPC, exploratory investigation of predictive biomarkers showed that B cell markers were among the most differentially expressed genes in the tumors of responders compared to non-responders and that TLS was associated with higher ORR [
45]. Results from a prospective study also showed that the presence of TLS was associated with better EFS [
79]. Using single-cell and spatial transcriptomic techniques, Liu
et al. identified pivotal cell populations associated with TLS that formed an immune-activated niche within NPC tissues, where germinal center reaction fostered the maturation of plasma cells that promoted the apoptosis of NPC cells, and CXCL13
+ cancer-associated fibroblasts promoted B cell adhesion and antibody production. The TLS-related cell signatures were found to correlate with immunotherapy response [
80]. Another study demonstrated the presence of a PD-1
+CXCR5
–CD4
+ Th-CXCL13 cell subset in NPC, which contributed to the production of CXCL13, recruitment of tumor-associated B cells and induction of plasma cell differentiation and immunoglobulin production [
81]. Lv
et al. showed that GP chemotherapy induced innate-like B cells (ILBs) via Toll-like receptor 9 signaling, which further expanded helper T cells and subsequently enhanced cytotoxic T cells in TLS-like structures that were deficient in germinal centers. Using a GP plus ICI cohort, a higher frequency of tumor-infiltrating ILBs was found to serve as a potential predictor of response to GP and ICI combined therapy and was significantly associated with longer PFS and OS [
82].
In recent years, other spatial organization within the TME have been discovered to have an association with response to anti-PD-1 immunotherapy. For instance, Chen
et al. described the stem-immunity hub, a kind of immunity hub distinct from mature TLSs, which was enriched for stem-like TCF7
+PD-1
+CD8
+ T cells, activated CCR7
+LAMP3
+ DCs and CCL19
+ fibroblasts as well as chemokines that organize these cells [
83]. Other studies emphasized the importance of cellular triads consisting of CD8
+ T cells, CD4
+ T cells, and DCs in mediating sufficient immunotherapy response [
84,
85]. With the development of spatial omics techniques like spatial transcriptomics, the existence of these structures and the spatial heterogeneity of the TME components along with their functions in mediating ICI response or resistance will be more comprehensively studied in NPC and other tumors, and whether they can serve as a reliable predictor of ICI benefit warrants further validation.
4.3 Multi-omics signatures
The characterization of genomic, epigenomic, transcriptomic, and proteomic alterations has remarkably improved the understanding of tumor biology of NPC and the TME. This holistic method has led to the identification of several useful prognostic biomarkers, which underscores the potential of multi-omics to discover predictive biomarker of ICI benefit and advance personalized medicine.
Several genomic alterations have been reported to predict ICI response. WES of tumor samples from patients with RM-NPC from two phase 1 clinical trials showed that copy number loss in either granzyme B or granzyme H (GZMB/H) genes was associated with poor survival after anti-PD-1 therapy [
86]. A phase 2, single-arm trial of nivolumab combined with gemcitabine in patients with RM-NPC found that the group with high somatic copy number alteration (SCNA) level had poor PFS [
39]. In the phase 2, single-arm POLARIS-02 study of toripalimab in patients with chemo-refractory metastatic NPC, genomic mutational analysis demonstrated that patients with genomic amplification in 11q13 region or ETV6 genomic alterations had poor responses to toripalimab [
36]. The 11q13.3 focal amplification along with the associated high expression of MAS-related GPR family member F (MRGPRF) was also reported in a pilot study to correlate with poor PFS from the triple combination of gemcitabine (chemotherapy) plus apatinib (anti-vascular endothelial growth factor (VEGFR)) and toripalimab (anti-PD-1) therapy (GAT) in RM-NPC patients [
87]. Wang
et al. developed a machine learning-based framework to explore extrachromosomal DNA (ecDNA) amplification, and showed that the circular subtype of focal copy number amplification (ecDNA+) was an independent prognostic factor for poor survival in NPC patients receiving PD-1 inhibitors [
88].
In addition, HLA-I heterozygosity was found to positively correlate with longer PFS after ICI therapy, potentially due to the increased diversity of MHC class I molecules on the cell surface, which enhances the presentation of a broader range of antigens recognized by cytotoxic T cells [
89]. Further, larger cohort studies are needed to determine whether strong correlations exist between genomic alterations and ICI efficacy in NPC patients.
Gene expression signatures also show potential to predict ICI responses. In a randomized phase 2 trial of spartalizumab (anti-PD-1) versus chemotherapy in RM-NPC, an exploratory analysis of RNA sequencing data demonstrated an association between the expression of immune activation and checkpoint molecules (IFN-γ, LAG-3, and TIM-3) and response in the spartalizumab arm, but not in the chemotherapy arm, indicating the potential predictive capability of these markers [
90]. Several other gene expression signatures have been proposed to predict ICI efficacy in NPC based on bioinformatics prediction, likely due to their association with immune infiltration and activation. For instance, annexin A6 (ANXA6) expression was shown to contribute to interleukin-2-mediated T cell proliferation [
91], and a high ferroptosis score positively correlated with CD8
+ effector T cell and antigen processing machinery [
92], both of which were associated with favorable ICI responses. In contrast, an RNA N
6-methyladenosine (m
6A) modification score was linked to negative regulation of immune checkpoint molecules and decreased ICI responses [
93,
94]. NPC-derived secreted phosphoprotein 1 (SPP1) promoted M2 macrophage via CD44/JAK2/STAT3 signaling and the SPP1-related M2 signature was predictive of inferior response to ICIs [
95]. However, validation of these signature in real-world cohorts is still warranted.
4.4 Molecular subtypes
Given the high dimensionality and complexity of omics data, a widely used approach is to perform clustering analysis of tumor samples to identify molecular subtypes with similar phenotypes and functions. These subtypes may exhibit distinct response patterns to different treatments, providing valuable insights for guiding precision medicine and optimizing treatment strategies. To this end, many molecular classifications of NPC have been established based on genomic or transcriptomic data [
96–
99], showing correlations between TME features and patient outcomes. However, due to the current lack of publicly available data of NPC patients receiving immunotherapy, most studies only relied on bioinformatics algorithms (e.g., Tumor Immune Dysfunction and Exclusion (TIDE) score [
100]) to predict ICI responses within each subtype or validated the subtypes in immunotherapy cohorts from other cancer types, without confirming the predictive capability of these subtypes specifically in NPC patients. Chen
et al. identified an immune-enriched subtype of NPC with abundant immune infiltration, which was further divided into evaded immune subtype (E-IS) and active immune subtype (A-IS). The E-IS subtype showed high expression of exclusion- and dysfunction-related signatures, such as TGF-β-associated extracellular matrix. In contrast, the A-IS subtype exhibited an activated immune profile and was associated with a favorable prognosis and improved response to immunotherapy in NPC patients [
101]. Whether this subtype can predict long-term survival benefit from immunotherapy remains further validation in prospective studies.
5 Tumor-related non-invasive biomarkers
Despite the advances in tumor tissue-based biomarkers, a major challenge in their clinical implementation for NPC is the limited availability of biopsy specimens, as surgical resection is rarely performed. Therefore, blood-based liquid biopsy and other non-invasive techniques such as radiological imaging are attracting growing interest for their ability to facilitate dynamic monitoring of disease progression and treatment response (Fig. 2).
5.1 EBV DNA
EBV DNA is considered one of the most promising biomarkers in NPC. As a distinct type of cancer closely associated with EBV infection, NPC cells release cell-free EBV-derived DNA fragments into the systemic circulation, which can be quantified by the quantitative reverse transcription polymerase chain reaction (qRT-PCR) method with probes against EBV genes [
102,
103]. The concentration of EBV DNA closely correlates with tumor progression and disease burden [
104]. This phenomenon establishes EBV DNA as a surrogate biomarker that not only reflects real-time tumor activity but also provides a non-invasive means to monitor treatment response and predict clinical outcomes in NPC patients. An increasing body of evidence supports the prognostic role of plasma EBV DNA in the management of NPC patients. Baseline and longitudinal evaluations of plasma EBV DNA during induction chemotherapy have been shown to be powerful prognostic factors for patients with locally advanced disease. In regard to immunotherapy, although some retrospective studies or single-arm trials proposed that baseline levels of plasma EBV DNA could be a useful biomarker for outcomes in patients with RM-NPC who received anti-PD-1 treatment [
35,
40,
105–
107], subgroup analyses from large-scale RCTs suggested improvement of survival in all baseline EBV DNA subgroups (Table 4).
Dynamic monitoring of plasma EBV DNA during the course of anti-PD-1 therapy has also shown potential as a predictor of treatment efficacy. A significant correlation between early reduction in plasma EBV DNA levels and better ORR or survival in patients with RM-NPC receiving anti-PD-1 therapy has been demonstrated in single-arm trials and real-world data [
35,
36,
108,
109]. However, some inconsistent results have been observed. In the post-hoc analysis of a phase 2, randomized controlled study, patients with undetectable plasma EBV DNA after neoadjuvant therapy showed a 3-year PFS of 100%, while PFS benefit from immunotherapy was observed in patients with detectable plasma EBV DNA after neoadjuvant therapy [
55]. The phase 2 NCI-9742 study of anti-PD-1 therapy in RM-NPC patients receiving at least one prior line of chemotherapy reported no association between survival and plasma EBV DNA clearance [
110], while in the randomized controlled CAPTAIN-1st study of first-line anti-PD-1 therapy in RM-NPC, early clearance of plasma EBV DNA was associated with prolonged PFS both in the ICI group and placebo group [
33], indicating that it may not serve as a biomarker to distinguish patients who can benefit from ICIs.
Nevertheless, these findings still highlight the importance of incorporating EBV DNA testing into the design of future clinical trials. Challenges that remain regarding the use of EBV DNA in immunotherapy for NPC include the standardization of EBV DNA detection methods. EBV DNA testing needs to be conducted in qualified central laboratories to ensure consistency and comparability of results. Additionally, previous studies have employed various cut-off values for EBV DNA plasma concentrations, but a standardized threshold has yet to be established. Standardizing methods for EBV DNA assessment and defining uniform cut-off values, along with further validation in future studies, are urgently needed.
5.2 Other liquid biopsy assays
Given the known sampling-related limitations with tissue-based analyses, the alternative use of blood-based analyses, known as liquid biopsy, is less invasive, easily accessible, potentially more cost-effective, and associated with less pain to the patient as well as a lower risk of complications. This approach also offers the advantage of biological assessments across all tumor lesions within an individual and is thus well positioned to assess dynamic treatment-related changes. The above-mentioned plasma EBV DNA detection is one of the most widely used liquid biopsy approaches for NPC patients. Besides, circulating tumor DNA (ctDNA) and circulating tumor cells are common liquid biopsy methods that have been used to evaluate tumor persistence and guide treatment decisions. These include determining the appropriate timing for initiating adjuvant immunotherapy in sensitive tumors and selecting patients who may benefit from additional treatments [
111].
With the aim of exploring the utility of ctDNA as a biomarker in patients with advanced cancers undergoing immunotherapy, Zhang
et al. conducted a pan-cancer analysis using pretreatment and on-treatment ctDNA samples across 16 types of advanced tumor from three phase 1/2 trials of durvalumab, pretreatment ctDNA levels appeared to be prognostic and on-treatment dynamics to be predictive. A molecular response metric based on the dynamics of variant allele frequency (VAF) was proposed to be potentially predictive of benefit from ICI treatment [
112], but only six NPC patients were included in this study. In the aforementioned pilot study on GAT therapy in RM-NPC, You
et al. observed that patients with negative ctDNA after two cycles of treatment showed significantly better PFS compared to those with positive ctDNA [
87]. However, the predictive value was not strictly validated due to the lack of controlled groups of patients who did not receive ICI treatment in these studies. Whether ctDNA can predict the efficacy of immunotherapy in NPC warrants further validation in larger cohorts.
5.3 Radiomics
The diagnosis and radiotherapy of NPC heavily rely on imaging modalities such as MRI and CT, generating a vast amount of imaging data. This wealth of data forms a robust foundation for the advancement of radiomics. Radiomics enables the extraction and analysis of high-dimensional quantitative features from medical images, offering valuable insights into tumor characteristics such as heterogeneity, shape, and texture that are often not visible to the human eye. These features can complement clinical and genomic data, contributing to improved diagnostic accuracy, and personalized therapeutic strategies for NPC [
113,
114]. Moreover, the integration of machine learning techniques with radiomics further amplifies its potential to identify predictive biomarkers and guide precision immunotherapy. In a retrospective study, Zeng
et al. developed a combined nomogram model, incorporating radiomics features obtained from pretreatment CT and clinical features, to predict treatment response and survival outcomes for LA-NPC patients receiving multiple types of induction chemotherapies, including immunotherapy and targeted therapy [
115]. A recent study evaluated a predictive model for the efficacy of PD-1 inhibitor combined with GP chemotherapy using MRI-derived deep learning features (DLFs). Random forest algorithm was employed to identify the most important features from 99 patients with advanced NPC. This MRI-based model effectively predicted the clinical complete response to PD-1 inhibitor combined with GP chemotherapy [
116]. Another study by Lin
et al. discovered distinct radiomics features between NPC patients with metastases emerging via the lymphatic route and those with metastases emerging via the hematogenous route, and a radiomics model was utilized to identify patients in the hematogenous group who had significantly better PFS and anti-PD-1 response [
117]. These findings highlight the potential clinical utility of radiomic models, offering a valuable tool for optimizing treatment management in NPC patients.
6 Biomarkers in the tumor macroenvironment
The aforementioned assays mainly offer information on tumor existence and therefore may need to be integrated with other indicators to measure immune activation or host immune competence. Cancer is now well recognized as a systemic disease. The effects of tumor-induced perturbations of the immune system extend beyond the local tumor immune microenvironment, resulting in the interactions between the TME and what is called the tumor macroenvironment (TMaE). Several alterations in the TMaE during tumorigenesis and development have been depicted, including systemic inflammatory alterations, systemic metabolic and microbiome changes [
118] (Fig. 2). These systemic cues can be detected using minimally invasive means such as a simple blood test, making them well positioned to be used as predictive biomarkers and to monitor treatment response. Deciphering the functional capacity and stability of the TMaE is critical for improving immunotherapy.
6.1 Immune and inflammatory indicators
Many promising discoveries have been made in regard to immune-related molecules in plasma or serum that can be detected to reflect systemic immune status. Advances in detection techniques have fostered the discovery of possible predictive biomarkers for immunotherapy. Lin
et al. reported a multiplexed enhanced fluorescence microarray immunoassay (eFMIA) that outperformed traditional fluorescence immunoassays (FIA) and enzyme-linked immunosorbent assays (ELISA). Through eFMIA, they found significantly differential levels of soluble PD-L1 (sPD-L1) and soluble intercellular adhesion molecule-1 (sICAM-1) in the sera of 28 cancer patients, with different clinical outcomes following anti-PD-1 therapy [
119]. Wu
et al. constructed a platform with high sensitivity to detect serum interleukin-15 (sIL-15) and conformed the predictive value of sIL-15 in two independent cohorts of 130 sera from patients with advanced NPC [
120]. The composition and dynamic change of peripheral blood mononuclear cells (PBMCs) can reflect the TME as well as systemic immune status. Huang
et al. performed peripheral blood immune profiling and demonstrated that the frequency of Ki67
+ Tregs was at a significantly higher baseline level and kept rising during anti-PD-1 therapy in LA-NPC patients who experienced relapse compared to those who did not. The predictive ability of baseline frequency of Ki67
+ Treg was further confirmed [
75].
Hematologic and inflammation-related parameters also matter and have been widely studied, including the neutrophil-lymphocyte ratio (NLR), the platelet-lymphocyte ratio (PLR), the lymphocyte-to-C-reactive protein ratio (LCR), and the lymphocyte-monocyte ratio (LMR). The systemic immune-inflammation index (SII) is defined as follows: SII = platelet count × neutrophil count/lymphocyte count [
121]. Lower baseline SII and higher baseline LMR were related to better PFS. The dynamic changes of SII and LMR were independent prognostic factors for the survival of NPC patients receiving ICIs [
122,
123]. Other studies constructed SII-based nomograms to accurately predict PFS of NPC patients receiving PD-1 inhibitor immunotherapy [
124,
125]. The lung immune prognostic index (LIPI) was developed based on derived neutrophils/(leukocytes minus neutrophils) ratio (dNLR) greater than 3 and lactate dehydrogenase (LDH) greater than upper limit of normal (ULN), characterizing 3 groups (good, 0 factors; intermediate, 1 factor; poor, 2 factors) [
126]. Pretreatment LIPI and its longitudinal variations may serve as potential biomarkers for predicting immune checkpoint inhibitor outcomes in RM-NPC patients [
127]. Taken together, these biomarkers serve as dynamic indicators of the immune–inflammatory balance. A favorable equilibrium characterized by attenuated immunosuppressive inflammatory signals coupled with potentiated cytotoxic T cell responses will ultimately enhance therapeutic efficacy and improve survival outcomes.
6.2 Systemic biochemical alterations
Systemic biological alterations, including the metabolic disturbances and the nutrition status, are also associated with immunotherapy response. Dysregulated lipid metabolism is a hallmark of cancer and may be reflected in the tumor macroenvironment. Higher baseline apolipoprotein A-I (ApoA-I) and high-density lipoprotein-cholesterol (HDL-C) levels, or an increase in ApoA-I and HDL-C after anti-PD-1 treatment could predict superior PFS in patients with NPC who received immunotherapy, an effect possibly associated with their ability to promote M1-like polarization of macrophages, thereby alleviating the immunosuppressive TME [
128,
129]. In terms of nutritional status, various indices such as the prognostic nutritional index (PNI) and the nutritional risk index (NRI) have been proposed, which are calculated based on body weight, the concentration of serum albumin, and total lymphocyte count. Guo
et al. constructed a nomogram using baseline nutritional (NRI, PNI) and inflammatory parameters (SII), along with uric acid and post-treatment EBV DNA that could improve the prognostic risk stratification for patients with metastatic NPC receiving chemotherapy plus anti-PD-1 treatment [
130].
Other factors include the organ dysfunctions. Patients developing thyroid dysfunction following combined IMRT and PD-1 inhibitor therapy demonstrated significantly improved overall survival [
131]. In a retrospective study of non-metastatic NPC patients with or without anti-PD-1 treatment, thyroid dysfunction was associated with better response to treatment in immunotherapy but not in standard treatment, suggesting its predictive value for survival benefit from additional ICI treatment and the need for regular monitoring of thyroid function in clinical practice of anti-PD-1 treatment for non-metastatic NPC [
132]. Liver function indicators were also shown to be able to predict survival for NPC patients who received anti-PD-1 immunotherapy. Baseline serum LDH level was identified as a potential determinant of ICI therapy outcome in RM-NPC [
133], and monitoring dynamic changes in LDH level and the ratio of aspartate aminotransferase (AST) level to alanine aminotransferase (ALT) level (AST/ALT) was shown to be predictive and prognostic for RM-NPC patients treated with PD-1 inhibitors [
134]. A nomogram on the basis of liver function-related indicators before ICI treatment was constructed to predict the efficacy of immunotherapy in NPC patients [
135].
Of note, the aforementioned parameters may be complementary to one another, and a clinically useful model may require careful selection and incorporation of these blood-based indicators. Wu
et al. implemented the least absolute shrinkage and selection operator (LASSO) Cox regression analysis to facilitate feature selection and the construction of a model to predict anti-PD-1 responses, which was composed of histologic subtypes, CD19
+ B cells, natural killer (NK) cells, regulatory T cells, red blood cells, AST/ALT ratio, apolipoprotein B, and LDH [
136]. Much effort has been made to conduct such model using machine-learning based methods in NPC or in the pan-cancer setting [
137,
138], and some of them are confirmed to be able to select patients who may benefit from the addition of anti-PD-1 treatment [
137].
6.3 Microbiota
Gut microbiota has been identified as a factor affecting the efficacy of immunotherapy across various cancers such as melanoma, lung and colon cancer. Evidence accumulates that modulating gut microbiota can affect both innate and adaptive immune responses during cancer therapy [
139], suggesting its relevance in NPC as well. In a study by Xu
et al., faecal samples from 57 advanced NPC patients receiving PD-1 inhibitors with nivolumab or camrelizumab were collected. The diversity of gut microbiota was not found to be associated with treatment response, but in the gut of non-responders to anti-PD-1 treatment, a higher baseline abundance of seven specific gut bacteria was found, including
Blautia wexlera,
Blautia obeum, four other bacteria belonging to the Clostridiales order, and one
Erysipelatoclostridium, emphasizing the intricate relationship between gut microbiota composition and immunotherapy outcomes [
89]. In another study by Yu
et al., faecal and blood samples from patients with advanced NPC were collected and subjected to 16S rDNA sequencing and survival analysis. The results showed that the abundance of
Lachnoclostridium and the alpha diversity were higher in the non-responding group than in the responding group. Patients with a lower abundance of
Lachnoclostridium had better PFS. Non-targeted metabolomics analysis revealed that
Lachnoclostridium affects the efficacy of immunotherapy through the usnic acid [
140].
In addition to gut microbiota, the polymorphic variability of microorganisms that symbiotically inhabit the body barrier exposed to the external environment, including the epidermis and the internal mucosa, and that can be detected within solid tumors, are considered to be one of the hallmarks of cancer [
141]. The bidirectional roles of intratumoral microbiota on cancer immunotherapy response is becoming evident, with several mechanisms involved. The intratumoral microbiota can promote cancer immunity through activation of cGAS/STING signaling, T and NK cell activation, maturation of TLS, and intratumoral microbiota-derived antigen presenting, while also dampen anti-tumor immunity through producing ROS, promoting an anti-inflammatory environment, and immunosuppression [
142]. Additionally, abundance of the identified intratumoral bacteria was associated with ICI response in some cancers, such as
Fusobacterium that was negatively associated with response to ICI in NSCLC [
143]. A recent retrospective cohort study including 802 NPC patients confirmed the existence of intratumoral microbiota in NPC, and high intratumoral bacterial load was found to correlate with poor disease-free survival [
144]. In the context of NPC, whether intratumoral microbiota load can serve as a predictor of immunotherapy efficacy remains further investigation and validation.
7 Challenges and perspectives
7.1 Deep multi-omics profiling to understand the TME
The sophisticated ecosystem of TME in NPC plays a pivotal role in tumor progression, immune evasion, and treatment resistance. Omics technologies bring unique advantages to the identification and validation of novel biomarkers at different biological levels. For instance, since NPC is not regarded as a mutation-driven malignancy, epigenetics and transcriptomics play a pivotal role in uncovering key oncogenic alterations and mechanisms of immune evasion [
145]. Additionally, given the abundant immune infiltration of NPC, state-of-the-art spatial transcriptomic and proteomic techniques offer powerful tools for visualizing and examining the intricate intercellular communication within the TME of NPC [
80,
146].
However, each single-omics technology alone is insufficient to fully comprehend the complex molecular landscape of tumors, necessitating the integration of multi-omics approaches. To this end, integrated multi-omics classification of cancer is an effective way to deal with the highly-dimensional and complicated data and facilitate the identification of robust predictive biomarker. Unfortunately, the clinical use of biomarkers in NPC is hindered by the lack of access to tissue as diagnostic specimens are often limited to nasopharyngeal endoscope and surgery is rarely performed, suggesting the need of alternative approaches for patient stratification with higher feasibility. In triple-negative breast cancer (TNBC), Jiang
et al. established four mRNA subtypes of TNBC and developed an IHC-based simplified approach for classification, allowing easy implementation and validation in prospective trials [
147–
150]. In small cell lung cancer (SCLC), Heeke
et al. showed that circulating-free DNA (cfDNA) methylation could be employed to distinguish transcriptional factor-defined SCLC subtypes [
151]. In urothelial carcinoma, four molecular subtypes with distinct TME and treatment benefit were identified, which could be predicted by digital pathology from H&E slides [
152]. These studies provide valuable pattern of discovering predictive biomarkers of ICI benefit in NPC.
7.2 Longitudinal monitoring and real-time adjustments
Immunotherapy profoundly alters the TME and the selective pressure fuels the emergence of resistant subclones, which, in turn, influence treatment response and long-term survival of patients. The development of minimally-invasive techniques reflecting TME-related or systemic information, such as EBV DNA and radiomics, can provide ongoing information about the genetic landscape and immune activity of tumor, enabling clinically feasible longitudinal monitoring of tumor progression and therapeutic response. Studying the dynamic changes in both the TME and the systemic immunity status is of great importance. On the one hand, though treatment-naïve baseline biomarkers do show promise in predict immunotherapy efficacy, patients within the same baseline biomarker subgroup will also behave distinctly during treatment, and the different response to treatment will also be a powerful biomarker that can predict long-term survival benefit from the treatment. On the other hand, longitudinal monitoring of these dynamic shifts in immune cell populations and the molecular landscape of the tumor is crucial for optimizing treatment strategies using adaptive designs to maintain therapeutic efficacy. Adaptive trial designs, where patient assignment to treatment arms is adjusted based on interim biomarker data, can further optimize predictive biomarker validation, allowing clinicians to adjust treatment protocols dynamically, switching between different ICIs or combining ICIs with targeted agents based on real-time data. Recent studies have depicted the longitudinal cell-free EBV DNA atlas during sequential chemoradiotherapy in NPC, which can inform real-time recurrence risk and promote risk-adapted, personalized patient treatment [
153,
154], providing a paradigm of research on immunotherapy in NPC. One of the main obstacles to the establishment of dynamic biomarkers lies in the optimization of detection timing and interval, which should be studied and confirmed in the future.
7.3 Incorporation of machine learning for predictive models
Recent advancements in medical technologies have led to the rapid accumulation of multidimensional data, including genomics, transcriptomics, proteomics, metabolomics, and other emerging omics data. Simultaneously, diverse forms of clinical data, such as medical records, imaging scans, and pathological images, have been explored and demonstrated to be prognostic or predictive, constituting the extensive multimodal data sets. Given the high dimensionality and complexity of these data, machine learning (ML) has shown distinct advantages in biomarker discovery, as it can effectively process intricate data sets and predict clinical disease progression through model construction and analysis.
One of the applications of ML to biomarker study is the establishment of multi-omics classification using unsupervised clustering algorithms. These methods can identify previously unrecognized patterns that are similar within each subtype but different across subtypes, with different outcomes after receiving immunotherapy. Supervised ML methods are also powerful to help with the identification of biomarkers that are most associated with treatment response or survival. Recently, various ML approaches have been developed for the discovery of treatment outcome-related biological features based on different types of data, such as CellCnn [
155] for mass cytometry data and Scissor [
156] for RNA sequencing data. Additionally, in order to deal with the multimodal data, deep learning based on neural network has been employed to generate predictive models, allowing for more tailored and effective treatment strategies. Moreover, this integrative approach not only improves model accuracy but also reduces the risk of overfitting through robust dimensionality reduction. In a study evaluating the advantages of multimodal data for survival prediction, multimodal models outperformed unimodal models in 24 out of 33 TCGA cancer types. The resulting models utilized all available modalities, highlighting the benefit of combining diverse omics data [
157]. Still, large-scale cohorts with data of high quality are needed to validate the robustness and generalizability of these models. In the era of precision medicine characterized by big-data-driven individualized treatments, effectively utilizing machine learning for biomarker discovery and rational implementation is both an opportunity and a challenge.
7.4 Contextualization in randomized controlled trials to validate predictive biomarkers
While numerous biomarkers have been proposed for their potential to predict responses to immunotherapy in NPC, without direct evidence from RCTs, it remains unclear whether these biomarkers genuinely predict the efficacy of a specific treatment or simply correlate with overall disease outcomes. To establish the predictive value of a biomarker, rigorous RCTs with appropriate control groups are essential. RCTs must compare outcomes between biomarker-defined patient subgroups receiving immunotherapy and a control group receiving standard treatment without immunotherapy (e.g., chemotherapy, radiotherapy, or targeted therapy). A biomarker-stratified design with sufficient sample size in each subgroup is ideal. However, if the prevalence of biomarker-negative cases is low, making a biomarker-specific design impractical, a biomarker-positive or overall approach may be justified. In such cases, biomarker-negative populations should still be analyzed as a secondary endpoint, and results must be thoroughly reported [
158].
Conducting biomarker-driven RCTs requires large patient cohorts and significant resources, especially for less common cancers like NPC. Collaborative efforts across multiple institutions and regions are crucial for ensuring sufficient statistical power and diversity in patient populations. Moreover, standardized protocols for biomarker assessment, such as uniform detection of PD-L1 expression or EBV DNA levels, are necessary to ensure reproducibility and reliability across studies. By designing RCTs that explicitly test the predictive value of biomarkers, researchers can distinguish between biomarkers that are truly indicative of immunotherapy benefit and those that simply reflect overall prognosis. This distinction is critical for guiding clinical decision-making and ensuring that patients receive the most effective, economical, and personalized treatments.
8 Conclusions
In conclusion, as immunotherapy continues to revolutionize the treatment patterns of NPC, challenges remain in optimizing personalized care under the guidance of predictive biomarkers. Much effort has been made to identify effective predictors of immunotherapy efficacy, resulting in the discovery of many tumor-based, blood-based and other non-invasive plausible predictive biomarkers. Multi-omics approaches, real-time monitoring of dynamic changes in the TME, integration with machine learning models, and robust validation through RCTs are key areas that will define the future of precision NPC immunotherapy. Addressing these challenges will ultimately enable clinicians to make more informed decisions, improving survival outcomes for patients with NPC.
PDF
(1664KB)
