Exploration of the benefits and resistance mechanisms of immunotherapy based on PD-L1 expression, circulating tumor DNA and T-cell receptor profiles in advanced gastric cancer.

Background
Gastric cancer (GC) is one of the leading causes of morbidity and mortality worldwide. China has a worldwide incidence of 50% incidence in the worldwide [1]. Most GC cases are diagnosed at advanced stages, which contributes toward poor prognosis [2]. In the era of immunotherapy, approaches for advanced or metastatic gastric cancer (AGC or MCG) have evolved into new times. CheckMate-649 [3], KEYNOTE-589 [4], ORENTIAL-16 [5], RATIONAL-305 [6], and GASTONE-303 [7] demonstrated superior outcomes of immune checkpoint inhibitors combined with chemotherapy to chemotherapy alone, especially for patients harboring programmed death ligand 1 (PD-L1) expression Combined Positive Score (CPS) > 5 in the human epidermal growth factor receptor 2 (HER2)-negative setting. Therefore, the National Comprehensive Cancer Network (NCCN) and European Society for Medical Oncology (ESMO) guidelines highly recommend the programmed death-1 (PD-1) inhibitor plus chemotherapy regimen in the first-line setting for HER-2 negative AGC or MGC if patients with PD-L1 CPS more than 5 benefit from these two landmark phrase III clinical trials (CheckMate-649, KEYNOTE-589). However, sintinimab and tisilimab have also granted regulatory approval for allcomers, regardless of PD-L1 status in China. Thus, the efficacy of immunotherapy in AGC or MGC patients with low PD-L1-expressing (PD-L1 expression CPS score < 5 or negative) remains controversial. One meta-analysis from the Journal of Clinical Oncology failed to demonstrate a survival advantage offered by the addition of ICIs to conventional chemotherapy in patients with low PD-L1-expressing tumors based on CheckMate-649, KEYNOTE-062, and KEYNOTE-590 [8]. These findings provide a basis for initiating a more detailed assessment of the selection of treatment options for first-line MGC. However, another meta-analysis that pooled PD-L1 CPS 1–4 subgroup Kaplan–Meier plots from KEYNOTE-859, CheckMate-649, and RATIONAL-305 showed a modest overall survival (OS) benefit with the addition of an anti-programmed cell death protein 1 (anti-PD-1) agent, including either MSI-H or excluding patients [9]. Thus, PD-L1 selected patients with AGC or MGC using PD-1 inhibitors are uncertain.
Furthermore, the survival benefit of combined immunotherapy approaches in patients with MGC differed, even in GC patients with identical clinical features. The mechanisms of intrinsic or acquired resistance warrant thorough investigation and management. Understanding these mechanisms is crucial for developing new combination treatment strategies to overcome drug resistance and enhance the efficacy of immunotherapy. The mechanisms leading to resistance to PD-1/PD-L1 inhibition are varied and multifactorial, involving multiple aspects such as intrinsic tumor factors, tumor microenvironment, host-related factors, and epigenetic regulation [10–13]. Several pan-cancer studies have reported that mismatch-repair deficiency, tumor mutation burden, microsatellite instability, and heterogeneity of neoantigens have a good predictive role in the efficacy of PD-1/PD-L1 inhibitors [14–16]. The above-mentioned biomarkers depend on tissue biopsy, but tissue biopsy is invasive and difficult to repeat, often failing to capture the genetic heterogeneity and temporal evolution of the tumor during the course of immunotherapy. In this regard, circulating tumor DNA (ctDNA) has garnered significant attention as a potential noninvasive biomarker [17, 18]. Concurrently, in AGC, ctDNA analysis has deciphered the molecular mechanisms underlying trastuzumab resistance through the identification of specific gene mutations within ctDNA. These findings lay the foundation for leveraging ctDNA to optimize cancer treatment strategies [19]. The dynamics of ctDNA by means of liquid biopsy after immunotherapy are easily obtained in clinical practice, and its predictive and prognostic value needs to be further confirmed. Previous studies have demonstrated that baseline characteristics and dynamic changes of ctDNA are directly correlated with the prognosis of patients with AGC and metastatic gastroesophageal cancer in the context of chemotherapy [20, 21]. In AGC immunotherapy, studies have demonstrated that patients with maximal somatic variant allelic frequency (maxVAF) of ctDNA reduction during treatment exhibit longer progression-free survival (PFS), OS, and higher response rates [22, 23]. However, the above studies were confounded by the inclusion of patients with multiline treatment or gastric cancer specimens with diverse HER mutation statuses. More homogeneous AGC cohorts should be included in future studies to validate the predictive role of ctDNA in immunotherapy.
T-cell receptor (TCR) detection has drawn increasing attention. TCRs exhibit extremely high specificity and diversity, which are mainly determined by the amino acid sequence of the hypervariable complementarity-determining region 3 (CDR3), and play a crucial role in anti-tumor immunity by specifically recognizing and binding to the tumor antigen-MHC complex [24, 25]. The diversity of the TCR repertoire (the total number of rearranged TCR sequences) enables it to bind to a wide variety of antigens [26]. Therefore, detecting the diversity of the TCR repertoire can reflect the diversity of the cellular immunity [27]. Next Generation Sequencing (NGS) technology allows rapid and accurate identification and quantification of the TCR repertoire [28]. Analysis of the TCR repertoire not only helps to understand the underlying immune mechanisms of tumorigenesis but also contributes to the exploration of new indicators for tumor diagnosis, treatment, and prognosis. A study characterized 374 clonally expanded tumor-specific CDR3s, whose roles in mediating antitumor immunity were highlighted in gastric cancer [29]. Moreover, in a phase Ib trial on neoadjuvant treatment in GC, it was demonstrated that neoantigen-specific T cell responses observed in all patients with pathological complete response (pCR) and clonal expansion of TCR correlated with systemic regression of tumor burden [30]. Existing studies on TCR profiling in AGC immunotherapy remain limited and warrant further investigation, particularly regarding the characterization of TCR signatures in immunotherapy-responding populations, which also constitutes a key component of the present study.
Thus, we conducted a retrospective study to analyze the efficacy of PD-L1 in selected AGC or metastatic patients and prospectively explore the dynamics of ctDNA and diversity of the TCR repertoire, aiming to explore features related to the efficacy of immunotherapy in HER-2 negative population.

Methods

Patients’ enrollment and sample collection
We retrospectively recruited HER-2 negative AGC or MGC who received first-line treatment between January 2018 and December 2024 at our hospital. Enrollment criteria: pathologically confirmed adenocarcinoma, locally advanced or metastatic (Stage IV), and cannot be surgically resected; the patient has received systemic first-line treatment (such as chemotherapy, targeted therapy, or immunotherapy) with complete records of the PD-L1 CPS, treatment regimen, dosage, cycles, etc., sufficient follow-up period, and clear records of progression-free survival (PFS). Exclusion criteria: Patients with pathology-confirmed non-adenocarcinoma type, lacking key treatment or survival data, and those who did not receive first-line treatment or whose first-line treatment was incomplete (the flowchart is shown in Fig. 1). This retrospective study was approved by the independent ethics committee of the Chinese People’s Liberation Army General Hospital (NO: S2019-201-01).
Peripheral blood samples were prospectively collected in Streek tubes from each patient at three time points: before the initiation of first-line immunotherapy (C1), after two cycles of treatment (C2), and during disease progression (P). Within 2 h, the samples were separated into plasma and peripheral blood lymphocytes (PBLs). Circulating cfDNA was extracted from the plasma using the QIAamp Circulating Nucleic Acid Kit (Qiagen). Germline genomic DNA was isolated from PBLs using the QIAamp DNA Blood Mini Kit (Qiagen, Hilden, Germany). Genomic DNA isolated from the PBLs was used for TCRβ sequencing. The workflow is shown in Fig. 3a.

Clinical efficacy assessments
Patients were followed up until the cut-off date of March 2025. Tumor evaluation was based on the Response Evaluation Criteria in Solid Tumors (RECIST, version 1.1) and included the objective response rate (ORR) and disease control rate (DCR). The main endpoint was PFS, which was defined as special treatment for RECIST-defined disease progression. The secondary endpoint was overall survival (OS), which was defined as special treatment for all-cause death.

NGS sequencing
Indexed next-generation sequencing (NGS) libraries were prepared from 80 ng of cfDNA and germline DNA using the NEBNext Ultra DNA Library Prep Kit for Illumina (NEB). The plasma and matched genomic germline DNA libraries were hybridized with a custom-designed 1021-gene panel covering a 1.5 Mb genomic region [31]. Sequencing was performed on a Gene + Seq-2000 sequencing system (Geneplus) with 100-bp paired-end reads.
The Burrows-Wheeler Aligner (BWA; version 0.7.12-r1039) was used to align the clean reads to the reference human genome (GRCh37.p13 (GCF_000001405.25)) after removing adaptors and low-quality reads. Picard (version 4.0, RRID:SCR_006525) was used to mark and remove duplicate reads. RealSeq (v3.1.0 Geneplus-Beijing, in-house) was used to filter duplicate reads using unique identifiers (UIDs) to remove errors generated by PCR or sequencing. Realignment and recalibration were performed using GATK version 3.4-46-gbc02625 (Broad Institute). DNAscope was used to call germline single-nucleotide variants (SNVs) and insertions and deletions (InDels). ApplyVarCal was used in the variant quality score recalibration (VQSR) stage.

Tumor somatic variant calling in ctDNA samples
Somatic SNVs and small InDels were profiled using GATK Mutect2 (version 4.1.4.1, RRID:SCR_001876) and realDcaller (v1.8.1; Geneplus-Beijing, in-house). Detailed variant calling and filter strategies have been reported previously [32]. Copy number variation (CNV) was analyzed using CONTRA (version 2.0.8, RRID:SCR_010814). The structural variation (SV) was analyzed using a self-developed NCsv (version 0.2.3 Geneplus-Beijing, in-house). All final candidate variants were manually verified using the Integrative Genomics Viewer (IGV).

ctDNA detection
After tumor somatic variant calling and annotation, mutations were filtered according to the following criteria: (i) no present in the matched control genomic DNA; (ii) no present in > 1% of the population in the 1,000 Genomes Project or dbSNP (RRID:SCR_002338) databases; (iii) its positional depth was at least > 300 × ; (iv) it was further filtered by an in-house list of recurrent artifacts based on a background estimate; (v) for hotspot mutations, ≥ 4 high-quality support reads, or for non-hotspots, at least ≥ 8 support reads; and (vi) clonal hematopoiesis was filtered through both backlist and sequencing results of paired WBC gDNA.

TCRβ sequencing
High-throughput sequencing of the CDR3β region was conducted using genomic DNA isolated from the PBLs. As described in a previous study, multiplex PCR primers were employed to amplify the rearranged TCRβ [33]. Subsequently, the constructed DNA library was sequenced on the Geneplus 2000 platform (Geneplus), with 100 bp paired‐end reads. To enable more in-depth bioinformatics analysis, downsampling to one million reads was carried out for each sample. Shannon entropy was calculated for the clonal abundance of all the productive TCR sequences. To determine the normalized Shannon entropy, the calculated Shannon entropy was divided by the natural logarithm of the number of unique productive TCR sequences, and this normalization process served as a proxy for assessing TCR diversity [34]. Clonality is defined as “1-normalized entropy” with values ranging from 0 to 1. To assess the similarity and overlap of TCR repertoires between different samples, the Morisita overlap index was utilized, which considers the specific T-cell rearrangements and their respective frequencies [35].

Statistical analysis
For survival data, Kaplan–Meier curves (log-rank test) were used to estimate the median OS and PFS and compare survival outcomes. The normality of the data distribution was evaluated prior to subsequent analysis. Exploratory univariate analyses were performed using the log-rank test, which compares different treatments and clinical variables. Multivariate analysis was performed using Cox regression analysis. Paired t-tests were used for paired-sample comparisons, and Wilcoxon tests, Fisher’s exact tests, and stratified k-square tests were used to compare the differences between the two groups. All P values were based on two-sided testing with statistically significant differences at P ≤ 0.05, and confidence intervals (CIs) were 95%. Statistical analyses were performed using the R software (v. 4.1.2) and SPSS software (18.0, RRID:SCR_002865). The ctDNA level (haploid genome equivalents per mL of plasma (hGE/mL)) was calculated by multiplying the mean VAF by the concentration of cfDNA (pg/mL of plasma) and dividing by 3.3, as previously described by Henriksen et al. [36].

Results

Patients’ baseline characteristics
Between January 2018 and December 2024, a total of 245 patients with advanced or metastatic gastric cancer (AGC or MCG) who received first-line treatment were finally enrolled in this retrospective study. A total of 170 (69.4%) patients were male and the median age of the cohort was 60 years (range, 21–82 years). The majority of patients (63.7%; 156/245) had a primary tumor located in the upper and middle parts of the stomach. According to pathological differentiation, 7.8% (19/245) were well-differentiated, 33.5% (82/245) were moderately differentiated, and 58.7% (144/245) were poorly differentiated. A total of 34.3% (44/255) of patients exhibited ≥ 2 metastatic lesions, with 38.0% (93/245) liver metastases and 37.6% (92/245) peritoneal metastases. Subsequently, to analyze the efficacy of PD-L1, the cohort was stratified by PD-L1 combined positive score (CPS): 33.5% (82/245) patients with CPS < 1, 25.7% (63/245) patients with CPS 1–4, and 40.8% patients (100/245) with CPS ≥ 5 (Fig. 2a). Significant intergroup differences were observed in age and number of metastatic organs (p < 0.05), while sex, tumor pathological differentiation, tumor primary location, and specific metastatic patterns (hepatic/peritoneal) showed no statistical difference (Table S1).

Survival and treatment efficacy analysis
Survival analysis and treatment efficacy were used to assess the effectiveness of immunotherapy in different PD-L1 expression subgroups. To avoid potential confounding by subsequent-line therapies, PFS, rather than OS, was selected as the primary endpoint, given that our investigation focuses on first-line treatment in patients with advanced gastric cancer. The median PFS (mPFS) for the overall cohort was 7.1 months (95%CI: 6.7–7.7 months). In the CPS < 1 subgroup, chemotherapy (C) alone demonstrated comparable PFS to combination therapies: C + PD-1 inhibitor (C + PD-1i) versus C + PD-1i + anti-angiogenic agent (mPFS 6.6 vs. 8.6 vs. 6.2 months, p > 0.05, Fig. 2b). In the CPS 1–4 subgroup, numerical differences in PFS were observed with C + PD-1i + anti-angiogenic therapy (mPFS 10.3 months) versus C + PD-1i (mPFS 7.4 months) or C alone (mPFS 7.5 months), but these differences did not reach statistical significance (p > 0.05, Fig. 2c). Notably, the CPS ≥ 5 subgroup confirmed established clinical patterns with improved PFS for C + PD-1i over C monotherapy (mPFS 6.9 vs. 5.8 months, HR 0.60, 95%CI 0.33–1.08, p = 0.040, Fig. 2d, Table S2). The OS was incomplete due to a loss of visit.
According to RECIST v1.1, objective responses were identified in 64 patients (0 complete response [CR] and 64 partial response [PR]), while 181 patients demonstrated non-response status (157 with stable disease [SD] and 24 with progressive disease [PD]) based on the best overall response (BOR) assessment. The ORR and DCR in the overall cohort were 26.1% (95%CI: 20.7–32.1%) and 90.2% (95%CI: 85.8–93.6%), respectively. With regard to tumor response in the subgroup, only the CPS ≥ 5 subgroup showed higher ORR in patients treated with C + PD-1i than in those treated with C alone (35.1% vs. 8.7%, p = 0.014, Table S3). These findings highlight the necessity of combination regimens of immunotherapy and chemotherapy in patients with a PD-L1 CPS ≥ 5.

ctDNA dynamics characterization
A longitudinal analysis was conducted on 55 plasma samples obtained from 20 patients (Fig. 3a). At the pretreatment time point (C1), 85.0% (17/20) of the patients showed somatic variations in the plasma samples. The most common mutation frequencies at C1 time point were TP53 (38%), LRP1B (19%), ARID1A (10%), CDH23 (10%), and EPHA3 (10%) (Fig. 3b). The cohort of 20 patients was stratified into long-PFS (n = 15) and short-PFS (n = 5) groups based on whether immunotherapy cycles were received (≥ 6 vs. < 6 cycles). Comparative genomics between the short-PFS and long-PFS groups demonstrated the temporal persistence of FAT2 mutations in the short-PFS group (100% detection across C1/C2/P timepoints vs. 20% in long-PFS, p = 0.008). In addition, a higher frequency of shared mutations across time points was observed in the short-PFS group (4/5) than in the long-PFS group (3/15) (Fig. 3b).
Early TP53/PREX2 variants (C1:80% vs. 13%, p = 0.004) and late TP53/NAV3 variants (P: 60% vs. 7%, p = 0.016) were detected in short-PFS and long-PFS groups, respectively (Fig. S1a, c). MYC amplification at C2 was exclusively observed in the short-PFS group (40% vs. 0%, p = 0.022, Fig. S1b). Quantitative analysis showed that mutational load did not intergroup significantly across timepoints (C1: p = 0.063; C2: p = 0.160; P: p = 0.430, Fig. 4a). In contrast, there was a significantly elevated ctDNA burden in the short-PFS group across all timepoints: maxVAF, C1 (p = 0.044)/C2 (p = 0.012)/P (p = 0.048); ctDNA level, C1 (p = 0.029)/C2 (p = 0.016)/P (p = 0.048) (Fig. 0.4b, c). However, we did not find a difference in the dynamics of the mutational load, maxVAF, and ctDNA levels between the groups (Fig. S2). Meanwhile, regarding the TOP mutations and mutations that exhibited significant intergroup differences, we discerned no dynamic differences across diverse time points (Fig. S3). Post-immunotherapy ctDNA clearance (defined as undetectable ctDNA levels) occurred in 50% (7/14) of long-PFS group versus 25% (1/4) in the short-PFS group (p = 0.032) (Fig. 4d, e). The longitudinal ctDNA profile suggests that sustained low ctDNA burden and effective ctDNA clearance may predict a durable immunotherapy response, whereas persistent/emerging mutations correlate with early progression.

TCR β-chain sequencing between short-PFS and long-PFS groups
To elucidate the baseline status of T-cell infiltration in different groups, the T cell repertoire in peripheral blood was characterized by sequencing the CDR3 region of the TCR β-chain in the short-PFS and long-PFS groups. We observed lower TCR clone counts in the short-PFS group than in the long-PFS group (p = 0.26, Fig. 5a). However, we did not find a significant difference in evenness (p = 0.69), Shannon index (p = 0.39), and clonality (p = 0.69) between groups, which were used to characterize the diversity of the TCR repertoire (Fig. 5b–d). The comparison of TCR distribution indicated that the T cells clones were similar between the long-PFS and short-PFS groups (Fig. 5e, f). When analyzing the shared clones between the two groups, we found that the distribution of shared clones did not clearly demarcate the long-PFS and short-PFS groups, resulting in a confounded clustering pattern (Fig. 5g). Although the two groups were relatively similar, we still identified some TCR clonotypes with significant differences in usage frequency that were enriched in the short-PFS and long-PFS groups respectively (Fig. 5h, i). These results provided preliminary evidence that the efficacy of immunotherapy might be related to specific CDR3 sequences, as opposed to a reliance on baseline TCR diversity; however, this observation remains tentative due to the current scope of data.

Discussion
Accurate identification of patients who can achieve long-term clinical benefit remains an unmet need in the field of gastric cancer immunotherapy. Although PD-L1 is a well-established biomarker in cancer immunotherapy, PD-L1 expression is known to exhibit significant spatial and temporal heterogeneity within the tumor microenvironment [37]. In addition, the correlation between PD-L1 expression and responsiveness to immunotherapy is relatively weak in gastric cancer. In ORENTIAL-16 study, subgroup of overall survival analysis showed HR was 0.90 (95%CI: 0.90–1.21) in patients with PD-L1 CPS < 5 [5] and 0.84 (95%CI: 0.51–1.37) in CPS < 1. A similar trend was observed in the CheckMate 649 study [3]. Meanwhile, in the FDA Oncology Advisory Committee Meeting launched on September 26, 2024 [38], the addition of ICIs to standard-of-care chemotherapy for the treatment of AGC or MCG patients with PD-L1 CPS < 1 does not appear to result in benefits. In our study, we further confirmed the non-superiority of PD-1 inhibitor in addition to chemotherapy when patients harboring the PD-L1 CPS < 5, even combining with anti-angiogenic agents, which is consistent with above studies. This is why many attempts have been made to combine other types of approaches, such as anti-VEGFR modality [39, 40] or other immune checkpoint inhibitors such as CTLA4 [41, 42], TIGIT [43], and LAG3 [44] inhibitors, aiming to improve the efficacy of PD-1 inhibitors. Technically, normalization of tumor vasculature can enhance tissue perfusion and improve immune effector cell infiltration, leading to immunotherapy potentiation [45, 46]. However, no clinical benefits were observed in the present study. In our clinical center, we conducted a phase II study of adding bevacizumab to anti-PD-1 and chemotherapy regimens in patients with PD-L1 CPS < 5 (the results have not been published yet, trial registration numbers: NCT05299476). A part of the population in this retrospective study was derived from this phase II study. However, we can see improved PFS and ORR in the CPS 1-4 group although there was no statistical difference in our study, consistent with a previous study [8]. Thus, the ongoing Phase III LEAP015 trial is expected to provide additional evidence about the role of anti-VEGFR tyrosine kinase inhibitors (TKIs) in combination with the standard of care.
The incidence of the PD-L1 CPS was not identical among the different regions. When we compared the PD-L1 status among CheckMate 649, KEYNOTE-585, and RATIONAL-305, the distribution of PD-L1 CPS < 1 and CPS 1-4 was 17%, 22%, 11%, and 22%, 29%, and 34%, respectively (Fig. 6) [38], respectively. In the ORENTIAL-16 study, Chinese patients with CPS < 5 accounted for 38.9% and CPS < 1 accounted for 15.4%. In our study, the proportion of patients with PD-L1 CPS < 1 was 33.5%, which was higher than that reported in other studies.
The longitudinal ctDNA analysis in our study revealed critical temporal patterns underlying immunotherapy outcomes in gastric cancer. The predominant pretreatment mutations (TP53: 38%, ARID1A: 10%) align with the gastric cancer genomic landscapes reported in TCGA-STAD [47]. Meanwhile, the temporal persistence of FAT2 mutations in the short-PFS group (100% vs. 20% in the long-PFS group, p = 0.008) suggested its potential role in mediating early resistance. Quantitative ctDNA analysis demonstrated significantly elevated maxVAF and ctDNA levels in the short-PFS group at all time points (Fig. 4b, c), indicating that the relative dominance of specific clones may drive therapeutic resistance. The higher ctDNA clearance rate in the long-PFS group (50% vs. 25%, p = 0.032) reinforces the role of ctDNA as a pharmacodynamic biomarker, consistent with the findings that molecular response assessment via ctDNA predicts clinical outcomes in immunotherapy trials [48]. The exclusive emergence of MYC amplifications at C2 time point in the short-PFS group (40% vs. 0%, p = 0.022) suggests that MYC-driven genomic instability may underlie immunotherapy resistance, as MYC could drive immune evasion via direct transcriptional upregulation of CD47 and PD-L1 [49]. These findings highlight the need for real-time ctDNA surveillance during immunotherapy. Other studies have also shown that dynamic changes in ctDNA can serve as potential biomarkers for predicting treatment efficacy and long-term outcomes in MGC patients [20–22].
TCR repertoire analysis revealed lower baseline clone counts in short-PFS patients (Fig. 5a), which may indicate that pre-existing expanded T-cell clones are prerequisites for a durable response [50]. However, the absence of significant diversity differences (Fig. 5c, d) was consistent with paradigms emphasizing clonal quality over quantity [51]. The lack of distinct clustering in shared TCR clones between groups (Fig. 5g) may reflect tumor microenvironment constraints or a limited cohort size. However, we still identified some TCR clonotypes enriched in either the long-PFS or short-PFS group, which may to some extent reflect that certain unique TCR clonotypes were associated with PFS.
There are two main limitations of our study regarding the analysis of ctDNA and TCR profiling. First, although we found an intriguing trend, the relatively small sample size introduced some inherent limitations. The use of a small sample size for the analysis of ctDNA and TCR must be considered exploratory, as it leads to insufficient statistical power and less robust conclusions. For instance, the non-significant difference in TCR diversity between the short and long PFS groups could likely be a consequence of the limited sample size. Consequently, larger cohorts are required in future research to verify the relationship between ctDNA/TCR monitoring and therapeutic outcomes in patients with AGC. Second, lack of post-baseline TCR data. Serial TCR sampling combined with mutation-specific neoantigen prediction may contribute to the dissection of clonal evolution dynamics during immunotherapy.

Supplementary Information
Below is the link to the electronic supplementary material.