Plasma immune proteome-based risk score predicts survival in advanced gastric cancer treated with PD-1 inhibitors and chemotherapy
Zhouwei Zhan, Ruyu Lin, Yigui Chen, Sha Huang, Luping Lin, Hanchen Zheng, Xiaojie Wang, Xiaoyan Lin, Zengqing Guo, Qiaoting Hu, Bijuan Chen
TL;DR
A blood-based protein test can predict survival in advanced stomach cancer patients treated with PD-1 inhibitors and chemotherapy.
Contribution
A five-protein plasma immune proteome risk score is developed to predict survival in gastric cancer patients undergoing chemoimmunotherapy.
Findings
A five-protein panel (LTB4R, GBP2, HLA-G, CYBB, HLA-B) was identified as a prognostic marker for survival.
The risk score outperformed traditional clinical and PD-L1 biomarkers in predicting overall survival.
A nomogram integrating the proteomic score with clinical variables provided accurate survival estimates.
Abstract
Blood-based biomarkers that capture systemic immunity could complement tissue-based assays for prognostication in advanced gastric cancer receiving programmed cell death protein 1 (PD-1)–based chemoimmunotherapy. We evaluated whether baseline plasma immune proteomics can stratify clinical outcomes and be operationalized into a clinically usable model. In a prospective cohort (n=40) treated with first-line PD-1 inhibitor plus chemotherapy, nano–ultra-high-performance liquid chromatography (nano-UHPLC) coupled with Orbitrap data-independent acquisition liquid chromatography–tandem mass spectrometry (DIA LC-MS/MS) was used to profile baseline plasma. Quality control (QC)-filtered protein intensities were median-normalized, log2-transformed, and batch-adjusted as needed. Group structure was assessed by principal component analysis (PCA). Differential expression (two-sided testing;…
Genes, proteins, chemicals, diseases, species, mutations and cell lines named across the full text — each resolved to its canonical identifier and authoritative record.
Click any figure to enlarge with its caption.
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6| Characteristic | Total (n=40) | STS (n=20) | LTS (n=20) | |
|---|---|---|---|---|
| Gender, n (%) | 0.197 | |||
| Male | 24 (60.0%) | 10 (50.0%) | 14 (70.0%) | |
| Female | 16 (40.0%) | 10 (50.0%) | 6 (30.0%) | |
| Age (years), n (%) | 0.057 | |||
| < 60 | 18 (45.0%) | 12 (60.0%) | 6 (30.0%) | |
| ≥ 60 | 22 (55.0%) | 8 (40.0%) | 14 (70.0%) | |
| ECOG PS, n (%) | 0.507 | |||
| 0 | 26 (65.0%) | 12 (60.0%) | 14 (70.0%) | |
| 1 | 14 (35.0%) | 8 (40.0%) | 6 (30.0%) | |
| Tumor differentiation, n (%) | 0.311 | |||
| High-Moderate | 13 (32.5%) | 5 (25.0%) | 8 (40.0%) | |
| Poor | 27 (67.5%) | 15 (75.0%) | 12 (60.0%) | |
| PD-L1 expression | 0.110 | |||
| PD-L1 positive (CPS ≥ 5) | 11 (27.5%) | 5 (25.0%) | 6 (30.0%) | |
| PD-L1 negative (CPS < 5) | 29 (72.5%) | 15 (75.0%) | 14 (70.0%) |
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsCancer Immunotherapy and Biomarkers · Ferroptosis and cancer prognosis · Gastric Cancer Management and Outcomes
Introduction
Gastric cancer (GC) remains a leading global health burden, ranking among the top causes of cancer death despite advances in multimodal therapy (1, 2). Immune checkpoint inhibitors (ICIs) have reshaped the treatment landscape, with approvals spanning first-line chemo-immunotherapy and later-line programmed death-1 (PD-1) monotherapy in selected subsets (3). Yet only a fraction of patients derives durable benefit, underscoring the need for robust, biology-anchored biomarkers to guide selection (4). Current clinic-ready markers, including programmed cell death 1 ligand 1 (PD-L1) expression, microsatellite instability-high/deficient mismatch repair (MSI-H/dMMR), Epstein–Barr virus (EBV) status, and human epidermal growth factor receptor 2 (HER2), capture important biology but suffer from assay variability, spatial/temporal heterogeneity, and imperfect positive predictive value in routine practice (5, 6). Beyond tumor-centric predictors, there is growing recognition that host immunity and systemic inflammatory tone critically condition ICI responsiveness in GC, motivating biomarker strategies that integrate tumor, microenvironmental, and systemic compartments (7).
Against this backdrop, proteomics has emerged as a compelling discovery and translation platform. Compared with genomics or transcriptomics, protein-level profiling directly reports effector pathways, post-translational regulation, and soluble immune mediators that shape antitumor immunity and resistance (8). In thoracic oncology, both tissue and blood proteomics have yielded signatures that stratify anti-PD-1/PD-L1 benefit more accurately than PD-L1 alone. One study used Sequential Window Acquisition of All Theoretical Mass Spectra (SWATH-MS) to quantify plasma proteins (9), while another integrated mass spectrometry (MS) with RNA sequencing and machine learning to analyze tumor tissue (10), each nominating immunobiologically coherent panels. Circulating proteome approaches are particularly attractive for gastrointestinal cancers: they are minimally invasive, repeatable across the treatment course, and capture systemic immune dynamics that tumor biopsies may miss. Recent high-precision plasma studies in immunotherapy-treated cohorts further reinforce the prognostic and predictive utility of immune-related protein panels, while enabling model integration with clinical variables for bedside use (11).
In advanced GC, translating these proteomic insights into clinically actionable tools remains a priority. Building upon evidence from cross-tumor proteomic studies and GC-specific reviews, we posit that pre-treatment peripheral blood proteomes can encode host-tumor immune interactions relevant to survival and early progression. A rigorously constructed, immune-informed protein signature could complement established clinicopathologic factors and PD-L1/MSI assessments, yielding a composite risk model with improved discrimination, calibration, and clinical interpretability. The present work therefore applies deep, high-coverage plasma proteomics in a newly diagnosed advanced GC cohort, defines immune-related differentially expressed proteins (DEPs), and uses regularized survival modeling to develop and evaluate a multi-protein risk score. Building on these insights, an immune-related proteomics-first strategy in advanced GC could yield minimally invasive risk models and mechanistic hypotheses, positioning plasma protein signatures as actionable tools to refine patient selection and to guide immunotherapy combinations.
Methods
Study design, participants, and ethics
This was a single-center, prospective observational study conducted at Fujian Cancer Hospital between October 2022 and August 2024. A total of 40 consecutive adults with histologically confirmed advanced GC were enrolled before initiation of systemic therapy. Eligibility criteria included age ≥18 years, Eastern Cooperative Oncology Group (ECOG) performance status (PS) of 0-1, adequate organ function, and no prior systemic therapy for advanced disease, and human epidermal growth factor receptor 2 (HER2)-negative status. HER2-positive patients were excluded from the study to avoid confounding from HER2-targeted therapies. Exclusion criteria comprised active autoimmune disease requiring immunosuppression, uncontrolled intercurrent illness or infection, and inability to comply with follow-up. All patients received first-line therapy consisting of a PD-1 inhibitor combined with chemotherapy (oxaliplatin 130 mg/m² on day 1 plus capecitabine 1,000 mg/m² twice daily on days 1–14 of a 21-day cycle), with dose modifications per Common Terminology Criteria for Adverse Events (CTCAE)-guided institutional protocol, administered every 3 weeks until disease progression, unacceptable toxicity, or withdrawal.
Baseline demographic and clinical variables collected included age, sex, ECOG performance status, tumor differentiation grade, and PD-L1 combined positive score (CPS). PD-L1 positivity was assessed using the 22C3 (or 28-8) antibody clone and Ventana detection kit, with positivity defined as CPS ≥5 (12, 13). Pretreatment peripheral blood samples were obtained within 14 days before therapy, processed under standardized protocols, and biobanked for proteomic profiling. The primary endpoint was overall survival (OS), defined as the interval from treatment initiation to death from any cause or last follow-up. The secondary endpoint was progression-free survival (PFS), defined as the interval from treatment initiation to radiographic or clinical progression or death. Tumor assessments followed institutional standards with reference to Response Evaluation Criteria in Solid Tumors (RECIST) v1.1 when applicable. Patients were classified into long-term (LTS) and short-term survivors (STS) using the cohort median OS as the cutoff. Follow-up was performed with imaging and clinical evaluations at predefined intervals until death or study completion on June 20, 2025. The study was approved by the Ethics Committee of Fujian Cancer Hospital, and written informed consent was obtained from all participants. All procedures adhered to the principles of the Declaration of Helsinki.
Sample collection and proteome preparation
Baseline peripheral blood was collected before first-line therapy into K2-ethylenediaminetetraacetic acid (K2-EDTA) tubes and processed within 2 h. Plasma was prepared by sequential centrifugation at 4°C, transferred to low-bind tubes, aliquoted to avoid repeat freeze–thaw, and stored at −80°C. Low-abundance proteins were enriched using a commercial magnetic-nanoparticle kit per manufacturer’s instructions: 20 µL bead suspension was magnetically separated and washed, resuspended, combined with 40 µL plasma, and incubated at 37°C for 30 min on a thermo-shaker. After magnetic separation, 150 µL wash solution was applied for 5 min and repeated three times. Captured proteins were reduced and alkylated, digested with trypsin to generate peptides, and desalted on C18. Total peptide concentration was measured by the nanodroplet method. Indexed retention-time (iRT) standards were added to the loading solvent immediately before analysis to support chromatographic alignment and run suitability. Pre-analytical variables (processing time, temperature, freeze–thaw policy) were standardized across samples; grossly hemolyzed specimens were excluded prior to digestion. The resulting peptide solutions were used for liquid chromatography–tandem mass spectrometry (LC-MS/MS) analysis.
LC-MS/MS acquisition and raw data processing
Proteomic data acquisition and primary analysis were performed by Shanghai OE Biotech Co., Ltd (Shanghai, China). Peptides were analyzed on a Thermo Scientific Vanquish Neo ultra-high-performance liquid chromatography (UHPLC) system coupled to an Orbitrap Astral mass spectrometer with an Easy-Spray source. Separation used a trap column (300 µm × 0.5 cm, 5 µm) and a C18 analytical column ES906 (PepMap™ Neo UHPLC, 150 µm × 15 cm, 2 µm). Mobile phase A was 0.1% formic acid (FA); mobile phase B was 0.1% FA in 80% acetonitrile (ACN). The gradient at 2.5 µL/min was: 0–4 min, 4–25% B; 4–5.8 min, 25–35% B; 5.8–6.2 min, 35–99% B; 6.2–6.9 min, 99% B. Data-independent acquisition (DIA) settings were: full MS 380–980 m/z at 240,000 resolution; automatic gain control (AGC) 500%; 2-Th precursor windows with 300 DIA windows; normalized collision energy 25%; tandem MS (MS/MS) 150–2000 m/z; radio-frequency (RF) lens 50%; maximum injection time 3 ms. Runs were block-randomized; pooled quality-control (QC) digests were injected every 5–10 samples, with blanks to monitor carry-over. Batch suitability was verified (spray stability, peak shape, iRT linearity, identification depth), and acquisition performance was trended (total ion current, mass accuracy, peptide/protein coefficients of variation [CVs]); non-conformant injections were reinjected or excluded by protocol. Raw DIA files were processed in DIA-NN (version 1.8.1). Parameters: enzyme trypsin (≤1 missed cleavage), variable oxidation (M) and N-terminal acetylation, fixed carbamidomethylation (C), false discovery rate (FDR) 1% at peptide-spectrum match (PSM) and protein levels. Protein intensities were exported, median-normalized, log_2_-transformed, and batch-adjusted as needed. Missing values were handled using a within-group left-censored approach. Sample-level QC (Pearson correlation, identification depth, protein-level relative standard deviation [RSD]) determined acceptance, yielding a harmonized matrix for differential and survival analyses.
Differential expression, functional enrichment, and immune focus
Log_2_-scaled protein intensities were compared between LTS and STS using two-sided tests with variance moderation where appropriate. Effect size was expressed as fold change (FC); DEPs were defined at FC ≥1.20 or ≤0.83 with p < 0.05 under Benjamini-Hochberg correction. Group separation and patterns were visualized by volcano plots and Z-score heatmaps with hierarchical clustering (Euclidean distance; Ward's linkage). Prior to hypothesis testing, principal component analysis (PCA) was performed on median-normalized, log_2_-transformed intensities (variance-filtered features) to assess global structure; 95% confidence ellipses summarized group dispersion, samples >3 SD from cohort centroids were flagged for QC review, and batch-adjusted PCA confirmed the outcome-associated separation observed in our results. Functional interpretation used over-representation analysis against the quantified background for Gene Ontology (GO) biological process (BP), cellular component (CC), and molecular function (MF) categories and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways with FDR control; enriched terms were prioritized by adjusted p-value and enrichment magnitude, and redundancy reduced via semantic-similarity filtering. To emphasize immune biology, Immunology Database and Analysis Portal (ImmPort) immune-related gene sets were mapped to protein identifiers, and immune-DEPs were defined as their intersection with DEPs while preserving directionality. Protein–protein associations among immune-DEPs were queried in Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) (Homo sapiens; confidence >0.7; text-mining edges down-weighted) and exported for network visualization and module inspection.
Survival modeling, validation, and reporting
Survival associations were first screened for immune-related DEPs (immune-DEPs) using univariate Cox proportional hazards models, reporting hazard ratios (HRs) with 95% confidence intervals and two-sided p values. Immune-DEPs then entered a least absolute shrinkage and selection operator (LASSO)-penalized Cox procedure (glmnet) with 10-fold cross-validation to select the penalty (λ-min) and retain features. The risk score was defined as the weighted sum of standardized abundances of the selected immune-DEPs, using their LASSO coefficients. Discrimination was evaluated by Kaplan-Meier curves, splitting patients at the cohort median risk score, with log-rank tests for statistical significance. Time-dependent receiver operating characteristic (ROC) curves and area under the curve (AUC) at 6, 12, 18, and 24 months quantified performance. Independence from clinical covariates (age, sex, grade, PD-L1 CPS) was tested in multivariable Cox models. A nomogram derived from the multivariable model provided 6/12/18-month OS estimates, with calibration assessed by bootstrap-corrected plots. Analyses were performed in R (4.4.1) with survival, glmnet, timeROC/survAUC, and rms; two-sided p < 0.05 was considered significant, with FDR control where applicable.
Results
Baseline characteristics and outcomes
Forty patients were enrolled and stratified by the cohort median OS into LTS (n=20) and STS (n=20). Baseline demographics and clinicopathologic features were broadly comparable between the STS and LTS group, including sex distribution, ECOG performance status, tumor differentiation grade, and PD-L1 CPS, supporting an unbiased discovery comparison (Table 1). The median follow-up for OS is 27.0 months. Kaplan-Meier analysis for the overall cohort showed a median PFS of 5.5 months (95%CI: 3.5-6.5 months) and a median OS of 10.0 months (95%CI: 7.1-10.9 months) (Figure 1).
Kaplan-Meier survival curves for the study cohort (n = 40). (A) Progression-free survival (PFS). (B) Overall survival (OS).
Global plasma proteomics and differential expression
DIA-LC-MS/MS profiling generated a high-coverage plasma proteome suitable for between-group comparisons. PCA demonstrated outcome-associated separation of LTS and STS along principal component (PC) 1 (PC1; 25.2%) and PC2 (6.82%), reflecting broad expression heterogeneity across survival groups (Figure 2A). Differential testing (p < 0.05 and |log_2_FC| > 0.263) identified 322 proteins, including 179 upregulated and 143 downregulated in LTS versus STS, visualized by volcano plots and a supervised heatmap (Figures 2B, D). Intersecting these proteins with curated immune catalogs delineated 36 immune-DEPs, alongside 285 non-immune DEPs and 1,775 immune proteins below the threshold, suggesting an immunologic contribution to survival differences (Figure 2C). Referenced Supplementary Table S1 for full details of the immune-DEPs and their associated statistics. Functional enrichment analysis highlighted adaptive and innate immune pathways, including complement activation (classical pathway), antigen processing and presentation, leukocyte-mediated immunity, and vesicular processes related to endocytosis and phagocytosis (Figure 2E). KEGG pathway analysis showed concordant enrichment, featuring natural killer cell-mediated cytotoxicity, nucleotide-binding oligomerization domain (NOD)-like and interleukin-17 (IL-17) signaling, hypoxia-inducible factor 1 (HIF-1) signaling, phagosome formation, and infection-related modules such as herpesvirus and Kaposi’s sarcoma-associated herpesvirus (KSHV), consistent with plasma-based host-tumor immune interactions (Figure 2F). Representative immune-DEPs included HLA-B, HLA-G, CYBB, GBP2, LTB4R, IFNGR1, and CD74, which co-varied with survival groups in the heatmap (Figure 2D).
Proteomic landscape, differential expression, and immune-related analyses based on DIA-LC-MS/MS (n = 40). (A) PCA of quantified proteins showing separation of LTS and STS. (B) Volcano plot of DEPs between LTS and STS, highlighting significant proteins at log2 fold-change thresholds and p < 0.05. (C) Venn diagram showing the overlap of DEPs with ImmPort-annotated immune proteins to define immune-DEPs. (D) Heatmap of representative DEPs with hierarchical clustering and group annotation, scaled by Z-scores. (E) Bubble plots of Gene Ontology (GO) enrichment in Biological Process, Cellular Component, and Molecular Function categories, with GeneRatio on the x-axis, adjusted p-value represented by color, and hit count by bubble size. (F) KEGG pathway enrichment dot plot of DEPs. DIA, data-independent acquisition; DEPs, differentially expressed proteins; immune-DEPs, immune-related DEPs; PCA, principal component analysis; LTS, long-term survivors; STS, short-term survivors.
Immune-DEP network and feature selection
To contextualize the immune signal, we mapped the immune-DEPs to the STRING interactome and observed densely connected modules enriched for antigen presentation, interferon signaling, and phagosome/complement pathways (Figure 3A). Proteins with higher degree and betweenness occupied module cores, suggesting potential coordination of innate–adaptive responses. Prognostic screening with univariate Cox regression identified a subset of immune-DEPs associated with OS and/or PFS, as visualized by ranked HRs and 95% confidence intervals (Figure 3B). We then applied LASSO-penalized Cox modeling with cross-validation to reduce collinearity and prevent overfitting, yielding a parsimonious multigene panel with stable selection across penalty paths (Figures 3C, D). The final multivariable Cox model retained five proteins (LTB4R, GBP2, HLA-G, CYBB, and HLA-B) with non-zero coefficients; effect sizes and confidence intervals are summarized in the forest plot (Figure 3E).
Interaction network, Cox regression, and LASSO modeling of immune-related candidate proteins (n = 40). (A) STRING protein–protein interaction network displaying evidence-weighted edges among candidate proteins. (B) Forest plot of HRs with 95% CIs from univariate Cox regression. (C) Ten-fold cross-validation curve for LASSO regression showing selection of the optimal λ-min. (D) Coefficient paths for candidate proteins, with features retained at the optimal λ-min. (E) Multivariable Cox regression forest plot of the final five-protein model, including overall model statistics. HR, hazard ratio; CI, confidence interval; λ, LASSO penalty parameter * p<0.05, *** p<0.001.
Construction and performance of the five-protein risk score
We constructed a prognostic risk score by fitting a multivariable Cox model on the five selected immune-DEPs (LTB4R, GBP2, HLA-G, CYBB, HLA-B) and computing a linear predictor as the weighted sum of Z-scored protein abundances. Using the cohort median as the cutoff, patients were dichotomized into high- and low-risk strata. The score distribution showed clear enrichment of higher values among patients with poorer outcomes, and event status aligned with increasing risk (Figures 4A, B). A supervised heat map confirmed coherent expression patterns of the five proteins across strata (Figure 4C). Kaplan-Meier analysis demonstrated marked separation for OS and PFS, with the high-risk group experiencing shorter survival; log-rank tests indicated statistically significant differences (Figures 4D, F). Time-dependent ROC curves further supported the score’s stability, showing consistent discrimination at 6, 12, 18, and 24 months for both OS and PFS (Figures 4E, G).
Stratification, expression, and prognostic discrimination of the five-protein risk model (n = 40). (A) Distribution of risk scores with median-based grouping into high- and low-risk categories. (B) Follow-up time and survival status by risk group. (C) Heatmap of Z-scored abundances of the five proteins annotated by risk group. (D) Kaplan-Meier curves for OS with numbers at risk shown. (E) Time-dependent ROC curves for OS at 6, 12, 18, and 24 months. (F) Kaplan-Meier curves for PFS. (G) Time-dependent ROC curves for PFS at 6, 12, 18, and 24 months. OS, overall survival; PFS, progression-free survival; ROC, receiver operating characteristic.
Comparison with clinical variables
Time-specific ROC analyses demonstrated that the five-protein risk score consistently outperformed conventional clinicopathologic covariates for OS discrimination across all assessed horizons (6/12/18/24 months). At the four time points, AUCs for the risk score were 0.850, 0.838, 0.911, and 0.844, exceeding those of age (0.416-0.511), sex (0.360-0.660), tumor differentiation grade (0.569-0.613), and PD-L1 CPS (0.425-0.722) (Figures 5A–D). Notably, PD-L1 CPS ranked second yet remained below the risk score at each time point (e.g., 0.679-0.722 vs. 0.844-0.911), indicating added value of the proteomic signature beyond established markers.
Time-specific ROC comparisons between the five-protein risk score and clinical variables for OS. (A–D) ROC curves at 6, 12, 18, and 24 months comparing the prognostic performance of the risk score with age, sex, tumor differentiation grade, and PD-L1 CPS. ROC, receiver operating characteristic; OS, overall survival; PD-L1 CPS, PD-L1 combined positive score.
Multivariable modeling, nomogram, and calibration
In Cox regression adjusted for age, sex, tumor differentiation grade, and PD-L1 CPS, the five-protein risk score remained independently associated with OS, with effect estimates that were directionally consistent with the univariate analysis (Figures 6A, B). Based on the multivariable model, we built a nomogram to estimate individualized 6-, 12-, and 18-month OS by integrating the risk score with clinical covariates (Figure 6C). Internal validation with bootstrap resampling demonstrated good apparent performance, and calibration plots showed close agreement between predicted and observed OS across all time points, with calibration curves aligning near the 45° reference line (Figures 6D–F). Collectively, these findings indicate that the proteomic risk score adds complementary prognostic information beyond routine clinicopathologic factors and can be operationalized into a bedside tool for survival estimation.
Prognostic evaluation integrating clinical variables with the five-protein risk score and nomogram calibration for OS. (A) Univariate Cox regression forest plot showing HRs, 95% CIs, and significance for individual variables (risk score, age, sex, tumor differentiation grade, PD-L1 CPS). (B) Multivariable Cox regression forest plot presenting independent prognostic effects within a combined model. (C) Nomogram derived from the multivariable Cox model; upper panels display point assignment for variables and total points, lower panels indicate predicted OS probabilities at 6, 12, and 18 months. (D–F) Calibration plots of the nomogram at 6, 12, and 18 months, where the x-axis shows predicted probabilities and the y-axis shows observed probabilities, with the 45° diagonal line representing perfect calibration. OS, overall survival; HR, hazard ratio; CI, confidence interval; PD-L1 CPS, PD-L1 combined positive score.
Discussion
In this proteomics-driven cohort of advanced GC treated with first-line PD-1 inhibitor plus chemotherapy, baseline plasma proteomes were shown to stratify patients by outcome, and an immune-enriched signature provided strong prognostic discrimination. Differential expression analysis revealed immune-related modules converging on antigen presentation and oxidative burst biology. Penalized feature selection identified a five-protein panel consisting of LTB4R, GBP2, HLA-G, CYBB, and HLA-B, and a weighted linear combination of these proteins produced a risk score that stratified OS and PFS with clear separation between high- and low-risk groups. Time-dependent ROC analysis demonstrated stable performance across 6 to 24 months, and multivariable modeling confirmed that the risk score contributed independent prognostic value beyond age, sex, grade, and PD-L1 CPS. A nomogram integrating the proteomic score with clinical covariates generated individualized survival estimates and exhibited good calibration. Taken together, these findings support a biologically coherent, blood-based tool that captures complementary information beyond conventional clinicopathologic features and PD-L1 CPS, offering a practical framework for risk stratification and individualized prognostication in the chemo-immunotherapy era for advanced GC.
Our findings align with a growing body of evidence that circulating immune proteomes can stratify benefit from ICIs across tumor types. Plasma-based signatures have predicted response and survival in triple-negative breast cancer, with immune and antigen-presentation modules enriched among top features (11), and multiplex plasma panels in non–small cell lung cancer (NSCLC) have similarly captured host-tumor crosstalk beyond tissue PD-L1, improving discrimination of clinical outcomes (14). In GC, tissue- and blood-derived markers of cytotoxic and stromal-immune activity have been described, but few studies have reduced complex profiles into compact, interpretable protein panels with prognostic value during first-line immunochemotherapy (15–17). Our five-protein panel, comprising LTB4R, GBP2, CYBB, HLA-B, and HLA-G, recapitulates pathways repeatedly highlighted in prior classifiers while maintaining accuracy with reduced dimensionality. Unlike NSCLC assays dominated by acute-phase and complement proteins (18), our model links antigen presentation with oxidative stress pathways, features especially relevant in GC’s inflammatory milieu. By integrating this score with routine clinicopathologic variables in a nomogram, we extend prior work on blood-based decision tools, demonstrating added value in a uniformly treated GC population receiving PD-1 plus chemotherapy.
Biologically, the five-protein signature encompasses complementary axes of antitumor immunity: LTB4R reflects leukotriene-mediated chemotaxis, GBP2 captures interferon-inducible resistance programs, CYBB encodes NADPH oxidase-driven oxidative burst, and HLA-B/HLA-G represent antigen presentation and immune tolerance (19–22). Their coordinated baseline variation suggests that systemic immune readiness in circulation influences benefit from PD-1–based therapy. This is consistent with proteomic studies in other cancers where immune and antigen-presentation modules predicted survival (11, 14). Clinically, a compact blood-based score provides several advantages over tissue biomarkers, as it is minimally invasive, repeatable, and integrates tumor, microenvironmental, and host inflammatory signals (23), which is particularly valuable in GC where biopsy samples are often constrained by fibrosis and heterogeneity. As a risk-stratification tool, the score could inform first-line triage: patients with high risk may merit intensified strategies (e.g., early incorporation of anti-angiogenic agents or clinical trials), while low-risk patients could avoid escalation beyond PD-1 plus chemotherapy. The calibrated nomogram further enables individualized survival estimation to guide counseling, imaging cadence, and palliative planning. Because the panel is mechanistically interpretable, it also generates hypotheses for combinations, such as pairing PD-1 blockade with therapies enhancing antigen presentation or reactive oxygen species (ROS)-mediated tumor killing, while complementing established markers like PD-L1 CPS and circulating tumor DNA (ctDNA) (24).
At the mechanistic level, the panel captures distinct but complementary immune processes measurable in blood. HLA-B reflects classical MHC-I–restricted antigen presentation supporting CD8+ T-cell priming, while HLA-G conveys a non-classical tolerogenic signal dampening cytotoxicity through ILT2/ILT4 (25, 26). GBP2, an interferon-inducible GTPase, indicates type I/II IFN activity and has been associated with antigen processing and inflamed tumor phenotypes (27). CYBB (NOX2) regulates phagocyte-derived reactive oxygen species, which can promote cross-presentation and immune activation but, when excessive, impair T-cell function (28). LTB4R coordinates eicosanoid-mediated leukocyte trafficking and T-cell infiltration, potentially shaping tumor immune landscapes under PD-1 therapy (29). Clinically, a plasma-derived score allows pretreatment risk stratification without tissue, complements PD-L1 CPS, and is easily serializable for monitoring. The nomogram illustrates how such a tool can guide counseling, surveillance, and trial enrollment. These implications mirror evidence that circulating immunoproteomic signals capture host-tumor interactions more broadly than single-tissue markers and can refine patient selection for immunochemotherapy across cancers.
This study has several strengths. First, baseline plasma was leveraged as a minimally invasive and clinically scalable source to derive an interpretable five-protein signature rooted in immune biology, including antigen presentation, interferon signaling, oxidative burst, and leukotriene chemotaxis. Second, uniform first-line PD-1 plus chemotherapy reduced treatment heterogeneity, while a transparent pipeline combining univariate screening, LASSO-penalized Cox regression, multivariable modeling, and nomogram calibration enhanced reproducibility. Third, the panel converges with recurring themes observed in blood-based proteomic studies of immunotherapy across cancers, thereby supporting external validity. However, limitations must be acknowledged. The single-center, modest sample size limits precision and introduces potential optimism bias from internally defined cutoffs, while the absence of external validation restricts immediate clinical application. Pre-analytical variability and platform transferability, such as DIA compared with targeted multiple reaction monitoring (MRM) or parallel reaction monitoring (PRM), were not tested. Similarly, data on MSI, EBV, tumor mutational burden (TMB), and ctDNA dynamics were incomplete, possibly underestimating integrative performance. Mechanistic inference also remains indirect because no functional assays verified causal roles of LTB4R, GBP2, CYBB, HLA-B, or HLA-G. Future research should prioritize prospective multicenter validation with locked models and prespecified thresholds, direct comparisons with PD-L1, TMB, and ctDNA, and health-economic studies assessing decision impact. Longitudinal sampling could clarify pharmacodynamic shifts under therapy, while experimental work is needed to define how leukotriene signaling, interferon-inducible pathways, ROS generation, and HLA-mediated antigen presentation shape treatment sensitivity or resistance, informing rational therapeutic combinations.
Conclusions
This study establishes a clinically oriented framework for leveraging circulating immune proteomics to inform prognosis and therapeutic planning in advanced GC treated with PD-1-based regimens. The resulting model provides an interpretable, blood-based signal that can be operationalized alongside routine clinicopathologic variables to support individualized decision-making and resource allocation. Its architecture is immediately adaptable to targeted assay formats and prospective workflows, enabling timely reporting within real-world care pathways. Future research should prioritize multicenter, prospective validation with locked algorithms and prespecified thresholds, rigorous evaluation against established biomarkers and ctDNA dynamics, and assessment of decision impact and health-economic value. Longitudinal sampling and pharmacodynamic modeling are warranted to map temporal immune trajectories under treatment and to guide adaptive strategies. Mechanistic studies dissecting leukotriene signaling, interferon-inducible programs, reactive-oxygen pathways, and HLA-mediated antigen presentation will be essential for designing rational combinations and for refining predictive enrichment. Collectively, these steps will accelerate translation toward equitable, guideline-ready implementation.
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1Morgan E Arnold M Camargo MC Gini A Kunzmann AT Matsuda T . Laversanne M et al: The current and future incidence and mortality of gastric cancer in 185 countries, 2020-40: A population-based modelling study. e Clinical Medicine. (2022) 47:101404. doi: 10.1016/j.eclinm.2022.101404, PMID: 35497064 PMC 9046108 · doi ↗ · pubmed ↗
- 2Zhan Z Chen B Zeng Y Huang R Yu J Guo Z . Long-term trends and projections of stomach cancer burden in China: Insights from the GBD 2021 study. Plo S One. (2025) 20:e 0320751. doi: 10.1371/journal.pone.0320751, PMID: 40198592 PMC 11978042 · doi ↗ · pubmed ↗
- 3Li K Zhang A Li X Zhang H Zhao L . Advances in clinical immunotherapy for gastric cancer. Biochim Biophys Acta Rev Cancer. (2021) 1876:188615. doi: 10.1016/j.bbcan.2021.188615, PMID: 34403771 · doi ↗ · pubmed ↗
- 4Luo D Zhou J Ruan S Zhang B Zhu H Que Y . Overcoming immunotherapy resistance in gastric cancer: insights into mechanisms and emerging strategies. Cell Death Dis. (2025) 16:75. doi: 10.1038/s 41419-025-07385-7, PMID: 39915459 PMC 11803115 · doi ↗ · pubmed ↗
- 5Li W Xu M Cheng M Wei J Zhu L Deng Y . Current advances and future directions for sensitizing gastric cancer to immune checkpoint inhibitors. Cancer Med. (2025) 14:e 71065. doi: 10.1002/cam 4.71065, PMID: 40686307 PMC 12277934 · doi ↗ · pubmed ↗
- 6Wan JY Zhang D Wu XH Yang H Dong C . Progress of immune checkpoint inhibitors in gastric cancer. World J Gastrointest Oncol. (2025) 17:109613. doi: 10.4251/wjgo.v 17.i 8.109613, PMID: 40837753 PMC 12362548 · doi ↗ · pubmed ↗
- 7Hiam-Galvez KJ Allen BM Spitzer MH . Systemic immunity in cancer. Nat Rev Cancer. (2021) 21:345–59. doi: 10.1038/s 41568-021-00347-z, PMID: 33837297 PMC 8034277 · doi ↗ · pubmed ↗
- 8Kirsher DY Chand S Phong A Nguyen B Szoke BG Ahadi S . Current landscape of plasma proteomics from technical innovations to biological insights and biomarker discovery. Commun Chem. (2025) 8:279. doi: 10.1038/s 42004-025-01665-1, PMID: 40999057 PMC 12462477 · doi ↗ · pubmed ↗