Dynamic programmed death-ligand 1 and multimodal biomarkers in gastric cancer immunotherapy

Abstract

Immune checkpoint inhibitors targeting programmed death-ligand 1 (PD-L1) and programmed death-1 (PD-1) are promising therapies for advanced gastric cancer (GC). However, challenges remain in identifying the best biomarkers to guide anti-PD-L1/PD-1 therapy. There are two competing paradigms for selecting immunotherapy biomarkers in clinical practice. One favors static biomarkers, such as Epstein-Barr virus positivity, microsatellite instability status, and the PD-L1 combined positive score, due to their track record in clinical trials. The other emerging view prefers dynamic tumor biomarkers, such as circulating tumor DNA, exosomal PD-L1, and immune cell profiling, to better reveal therapy-induced immune changes. This article discusses three main questions for real-time assessment of anti-PD-L1/PD-1 therapy in advanced GC: (1) What complementary approaches can overcome static PD-L1 assessment limitations; (2) How dynamic tumor biomarkers could be used in clinical practice to improve monitoring and patient selection; and (3) How multimodal artificial intelligence could support personalized immunotherapy. These strategies may improve patient selection and enhance the clinical impact of immunotherapy in advanced GC.

Key Words: Gastric cancer; Programmed death-ligand 1 expression; Tumor biomarkers; Artificial intelligence; Immune checkpoint inhibitors

Core Tip: Emerging non-invasive, real-time biomarkers such as serial exosomal programmed death-ligand 1 (PD-L1), circulating tumor DNA, and circulating immune cells may help overcome the limitations of static PD-L1 assessment, thereby improving the efficacy of anti-programmed death-1/PD-L1 therapy in advanced gastric cancer. Validation of rapid, cost-effective, and point-of-care dynamic tumor biomarkers is essential before their integration into routine clinical practice. Moreover, integrating histopathological and radiological findings with standard tumor biomarker data through artificial intelligence offers the potential to overcome the main limitations of circulating tumor DNA next-generation sequencing and PD-L1 immunohistochemistry assessment.

INTRODUCTION

Gastric cancer (GC) remains a leading cause of cancer-related mortality worldwide, particularly in advanced stages, which pose significant therapeutic challenges[1,2]. Immune checkpoint inhibitors (ICIs) targeting programmed death-ligand 1 (PD-L1) and programmed death-1 (PD-1) have demonstrated improved objective response rates and overall survival in patients with advanced GC[3-7]. Despite these advancements, identifying optimal biomarkers for real-time monitoring of PD-1 inhibitor response in advanced GC remains an unresolved challenge and an ongoing topic of debate. Two competing paradigms exist for selecting immunotherapy biomarkers in clinical trials. One favors reliance on conventional and static tests due to their accessibility in routine. Notably, a recent large randomized trial reported a significant decline in serum levels of conventional biomarkers, including carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 72-4 (CA72-4), and carcinoembryonic antigen (CEA), after 12 weeks of sintilimab plus chemotherapy in patients with advanced GC[8]. However, growing evidence suggests that these biomarkers lack validated predictive value for response to PD-1 inhibitors[9-11], largely due to their inability to capture dynamic immune changes within the tumor microenvironment (TME), as well as their variable sensitivity and specificity[12-14]. Similarly, established static biomarkers, such as Epstein-Barr virus (EBV) positivity, microsatellite instability (MSI) status, and PD-L1 combined positive score (CPS), are routinely used to guide immunotherapy decisions in GC[15-17]. Yet, their clinical value in the monitoring of PD-1 inhibitor response remains controversial. One school of thought supports the continued reliance on these biomarkers, given their established role in clinical trials[18]. Conversely, an emerging perspective argues that single-time-point biopsies are inherently insufficient, as they fail to capture tumor heterogeneity, sampling bias, temporal evolution, and therapy-induced changes in TME[19-22]. This article addresses three burning key questions regarding the real-time evaluation of anti-PD-1/PD-L1 therapy in advanced GC.

WHICH COMPLEMENTARY APPROACHES CAN OVERCOME THE LIMITATIONS OF STATIC PD-L1 ASSESSMENT?

The challenge of monitoring PD-1 inhibitor response in advanced GC can be seen in two parts: Identifying optimal biomarkers to guide real-time assessment of treatment response, and identifying cost-effective, non-invasive biomarkers suitable for longitudinal monitoring. Several studies have focused on developing multi-regional tissue biopsies, spatial transcriptomics, single-cell RNA sequencing, and multi-omics analysis to better characterize the heterogeneity of PD-L1 expression in advanced GC[23-25]. These are important advances, but their invasive nature and limited feasibility for longitudinal monitoring restrict clinical utility. Thus, non-invasive liquid biopsy biomarkers have emerged as a more practical and dynamic alternative for monitoring responses to PD-L1 inhibitors in GC[26-28].

Unlike static PD-L1 assessment, serial exosomal PD-L1 (exo-PD-L1) provides real-time insight into immunotherapy response in GC. Increased circulating exo-PD-L1 is frequently associated with resistance to anti-PD-1/PD-L1 therapy across multiple malignancies[29,30]. Although biologically compelling, exo-PD-L1 remains investigational in advanced GC[31-33]. It is typically measured by isolating exosomes and detecting PD-L1 using enzyme-linked immunosorbent assay, Western blotting, or flow cytometry[34,35]. Nevertheless, these immunoassays lack full standardization and remain challenging to interpret clinically[36]. Further studies are required to establish cost-effectiveness and standardization for routine use.

Additionally, circulating tumor DNA (ctDNA) assessed by next-generation sequencing (NGS) from liquid biopsy can predict clinical benefit from anti-PD-1/PD-L1 therapy in human epidermal growth factor receptor 2-negative GC patients[37]. The ctDNA kinetics can detect molecular progression before radiographic changes become apparent[38]. However, the adoption of ctDNA NGS technology in clinical practice remains limited by high costs and long turnaround times (often approximately 15 days)[39]. As a result, its feasibility depends on the socioeconomic status of the patients. Moreover, the prolonged turnaround times of NGS may negatively affect patient outcomes, potentially increasing mortality[40].

Measurements of circulating immune cells may further provide insight into immunotherapy response in GC. While high CD8+ T-lymphocyte infiltration predicts a better response to sintilimab[41], repeated invasive biopsies carry risks for patients. PD-L1 expression on circulating immune cells enables real-time assessment of treatment response and resistance[42]. Furthermore, multiplex immunofluorescence, combined with digital and machine-learning approaches, may enable more precise characterization of TME signatures, thereby supporting personalized therapy selection[43,44]. Thus, these findings support a paradigm shift from static, tissue-based PD-L1 assessment toward dynamic and multimodal biomarker strategies. Such an approach is likely to significantly enhance the efficacy of anti-PD-1/PD-L1 therapy in advanced GC by enabling real-time, personalized treatment adaptation.

HOW CAN DYNAMIC TUMOR BIOMARKERS BE INTEGRATED INTO CLINICAL PRACTICE?

Integrating dynamic tumor biomarkers into clinical practice requires a shift from static, single-time-point assessment toward a longitudinal, multimodal framework. Thus, a potential approach may combine baseline molecular stratification (e.g., PD-L1, MSI status, EBV status, tumor mutational burden), serial liquid biopsy monitoring (e.g., circulating T cells, ctDNA, exo-PD-L1), lactate dynamics, and computed tomography (CT) imaging to optimize treatment selection and response evaluation in advanced GC[15,38,45-47]. Nevertheless, the optimal thresholds, timing, and clinical algorithms for these biomarkers remain unstandardized in GC[48]. Table 1 shows the proposed timeline for dynamic biomarker monitoring of treatment with PD-1 inhibitors plus chemotherapy in advanced GC. Measurement of baseline pre-treatment biomarkers, including PD-L1 CPS, MSI status, and EBV status, can guide early monitoring of PD-1 inhibitor response, particularly in centers with limited resources[49,50]. After approximately 6-8 weeks of treatment, ctDNA kinetics can help predict the objective response and survival outcomes. The first CT imaging, in combination with conventional tumor biomarkers (CEA, CA19-9, CA72-4), can be performed after 8-12 weeks of treatment for standard efficacy assessment[51]. Thereafter, ctDNA assessment every 12 weeks may detect molecular progression earlier than CT imaging. Thus, by capturing both temporal and spatial tumor heterogeneity, this strategy has the potential to refine predictive models in patients with advanced GC (Table 1). Additional prospective studies may help standardize dynamic PD-L1 protocols and enable their transition from research settings to routine clinical practice.

Table 1 Proposed dynamic biomarker monitoring in gastric cancer immunotherapy¹.

Ref.	Timepoint	Assessments	Evidence-based rationale
Jiang et al[38], 2024; Li et al[59], 2025; Jiang et al[60], 2023	Baseline (pre-treatment)	PD-L1 CPS, MSI status (MMR), EBV status, ctDNA (tumor-informed if possible), CEA/CA19-9/CA72-4	(1) PD-L1 CPS, MSI-H/dMMR, and EBV positivity are validated predictive biomarkers for ICI benefit in advanced GC; (2) Baseline ctDNA reflects tumor burden and genomic landscape; tumor-informed assays may offer improved sensitivity compared with tumor-naïve panels; and (3) CEA/CA19-9/CA72-4 are prognostic but not validated predictive markers for ICI response
Jiang et al[38], 2024; Jin et al[45], 2020	After 2 cycles (approximately 6-8 weeks)	ctDNA kinetics, CEA/CA19-9/CA72-4	Early decline or clearance of ctDNA correlates with objective response and improved survival in GC treated with ICIs ± chemotherapy; CEA/CA19-9/CA72-4 decline may reflect treatment response but lacks immunotherapy-specific validation
Jiang et al[61], 2023	First radiographic evaluation (approximately 8-12 weeks)	CT imaging (RECIST/iRECIST), tumor markers (CEA, CA19-9, CA72-4)	Standard efficacy assessment; follow-up every 8-12 weeks is recommended per iRECIST guidelines
Nguyen et al[62], 2025	Every 12 weeks thereafter	ctDNA monitoring (if available), imaging as per protocol	Serial ctDNA kinetics may detect molecular progression earlier than imaging; timing, thresholds, and clinical algorithms remain unstandardized in GC
Fan et al[32], 2019; Ayala-Mar et al[33], 2021	Exploratory (research setting only)	Exo-PD-L1	Preliminary GC studies suggest a potential association with disease progression; currently investigational with no standardized clinical application

¹Proposed monitoring intervals are derived from current literature and clinical practice patterns but are not yet standardized in gastric cancer immunotherapy.

GC: Gastric cancer; PD-L1: Programmed death-ligand 1; iRECIST: Immune response evaluation criteria in solid tumors; CPS: Combined positive score; MSI: Microsatellite instability; MMR: Mismatch repair; EBV: Epstein-Barr virus; ctDNA: Circulating tumor DNA; CEA: Carcinoembryonic antigen; CA19-9: Carbohydrate antigen 19-9; CA72-4: Carbohydrate antigen 72-4; dMMR: Deficient mismatch repair; ICI: Immune checkpoint inhibitor; CT: Computed tomography; Exo-PD-L1: Exosome-programmed death-ligand 1.

HOW CAN MULTIMODAL ARTIFICIAL INTELLIGENCE-BASED BIOMARKERS SUPPORT PERSONALIZED IMMUNOTHERAPY SELECTION?

Artificial intelligence (AI)-based biomarkers, enabled by deep learning and multimodal integration of clinical, radiological, histopathological, and biological data, represents a promising yet relatively understudied type of biomarkers for monitoring ICIs in GC[52-56]. This approach goes beyond traditional single-biomarker assessments, offering the potential to address the main limitations of ctDNA NGS and PD-L1 immunohistochemistry evaluation (Figure 1).

Figure 1 Multimodal artificial intelligence-based solutions for predicting response to programmed death-ligand 1 inhibitors in advanced gastric cancer. This figure illustrates multimodal artificial intelligence-based solutions for predicting personalized responses to programmed death-ligand 1 inhibitors in advanced gastric cancer. The framework integrates diverse data types, including circulating immune cell populations, standard biomarkers (programmed death-ligand 1, combined positive score, and microsatellite instability status), computed tomography-based radiomics, and whole-slide images. PD-L1: Programmed death-ligand 1; CPS: Combined positive score; MSI: Microsatellite instability; EBV: Epstein-Barr virus; WSI: Whole-slide images; CT: Computed tomography.

A successful application is the prediction of PD-L1 expression, MSI status, and immunotherapy response in GC using radiomics-based machine learning models derived from CT images[46,57-59]. Similarly, several studies have demonstrated the ability of deep learning systems to predict PD-L1 CPS quantification from whole-slide images[54,60-62]. Despite their exciting potential, the predictive performance of these models for response to PD-1 inhibitors remains limited to proof-of-concept retrospective studies with relatively small sample sizes. Furthermore, AI-based biomarkers demonstrate moderate discriminatory ability in predicting anti-PD-1 efficacy in patients with GC, with areas under the curve (AUCs) generally ranging from 0.70 to 0.83[46,57,63,64]. Moving forward, we advocate for the development and regulatory approval of multimodal AI-based solutions for PD-L1 testing in accordance with clear guidelines. Furthermore, these models should demonstrate acceptable analytical performance in biomarker prediction without adding demands beyond standard workflows.

CONCLUSION

The clinical integration of dynamic PD-L1 panels offers a promising avenue to refine anti-PD-1/PD-L1 therapy decisions in advanced GC. In parallel, the implementation of AI-based PD-L1 tools designed to address the main limitations of ctDNA NGS and PD-L1 immunohistochemistry measurement has yielded additional powerful tools. We therefore propose that these biomarkers undergo large-scale prospective validation of the cost-effectiveness, particularly in smaller care centers with scarce resources or in low-income countries. Additionally, special care should be taken at every step of integrating these biomarkers into clinical practice to minimize clinician workload.