DNA methylation profiles predicting response to anti-PD-1-based treatment in patients with advanced gastric cancer

Abstract

Background

Immunotherapy has shown promise in treating gastric cancer (GC), yet predicting its efficacy remains challenging. Here we investigated DNA methylation as a predictive marker for response of anti-PD-1-based treatment in GC.

Methods

A total of 99 GC patients treated with first-line anti-PD-1-based treatment were enrolled. In the model construction phase, 30 samples were analyzed using the Infinium MethylationEPIC BeadChip (850 K array) and 41 samples using Targeted Bisulfite Sequencing (TBS). Support Vector Machine-Recursive Feature Elimination (SVM-RFE) and Least Absolute Shrinkage and Selector Operation (LASSO) were applied to identify differential CpG methylation probes (DMPs). Seven machine learning models were developed, and their performance was assessed by the area under the curve (AUC) of receiver operating characteristic and survival analysis. SHapley Additive exPlanations (SHAP) analysis provided interpretability of the model. In the model validation phase, a temporally independent cohort of 28 samples underwent TBS for external validation.

Results

The 850 K array identified 523 DMPs, of which 20 were selected as most significant for treatment response. The iMETH model, based on the k-nearest neighbors (KNN) algorithm, showed optimal predictive value in both training (AUC = 0.99) and testing (AUC = 0.96) sets. Progression-free survival (PFS) and overall survival (OS) were significantly longer in responders predicted by iMETH (all log-rank test p < 0.05). SHAP identified cg06692537 as one of the most important features, indicating that hypomethylation of this probe was associated with a likelihood of benefiting from anti-PD-1 based therapy. The model’s robustness was confirmed in the validation set (n = 28, AUC = 0.83). 

Conclusions

Our findings demonstrate that DNA methylation markers can serve as valuable predictors of first-line immunotherapy in GC. The iMETH model, which incorporates 20 DNA methylation CpG probes, effectively predicts patient responses to first-line anti-PD-1-based treatment. This model holds promise for guiding personalized treatment strategies and has the potential for practical application in clinical settings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12916-025-04357-8.

Background

Gastric cancer (GC) is one of the most prevalent malignant tumors, characterized by high incidence and mortality rates globally [1, 2]. Systemic therapy plays a pivotal role in treating advanced GC, with immunotherapy targeting programmed death 1/programmed death ligand 1 (PD-1/PD-L1) showing promising outcomes [3]. Clinical evidence has highlighted significant benefits for certain patients. Landmark studies, including CheckMate-649 and KEYNOTE-859, have demonstrated the survival advantages of anti-PD-1 therapy in advanced GC patients [4]. More recently, the ORIENT-16 trial reported an objective response rate (ORR) of 48.4% with the combination of anti-PD-1 therapy and chemotherapy, extending median overall survival by 2.9 months compared to chemotherapy alone (15.2 months vs. 12.3 months) [5]. As a result, the combination of anti-PD-1 drugs and chemotherapy has become an important treatment option for advanced GC, particularly for patients who lack therapeutic targets such as human epidermal growth factor receptor 2 (HER-2) [6–8]. However, in clinical practice, a significant proportion of GC patients do not benefit from this therapeutic approach, instead facing substantial side effects and financial burdens [9–11]. This underscores the need for precise identification of patients who are most likely to benefit from first-line anti-PD-1-based treatment. 

Biomarkers such as PD-L1 expression, microsatellite instability (MSI), tumor mutation burden (TMB), and Epstein-Barr virus (EBV) infection status have been utilized to predict responses to immunotherapy in GC patients [12–14]. However, these markers alone fail to fully account for the variability in immunotherapy outcomes among GC patients. For example, a significant proportion of GC patients with low or negative PD-L1 expression can still benefit from anti-PD-1 treatment [15–17]. Additionally, GC patients with microsatellite instability-high (MSI-H) or EBV-positive status constitute less than 10% of cases [18, 19], making these markers lack representativeness in the broader patient population. Furthermore, high cost and lack of standardization limit the clinical application of TMB detection [20, 21]. Clearly, the predictive value of these biomarkers is insufficient for precise treatment decisions, highlighting the urgent need for more reliable markers.

DNA methylation is a prevalent epigenetic modification that plays a crucial role in maintaining genomic stability, transmitting genetic information, and regulating gene expression [22, 23]. Research has demonstrated that aberrant DNA methylation is commonly observed in various cancers, including colorectal, breast, and lung cancers, and is closely linked to tumor initiation, progression, and metastasis [24, 25]. These epigenetic alterations can function as diagnostic markers, prognostic indicators, and predictors of therapeutic response [26, 27]. Emerging evidence suggests that DNA methylation holds promise as a predictive biomarker for immunotherapeutic outcomes in several solid tumor types, such as lung cancer, head and neck squamous cell carcinoma, melanoma, and sarcoma [28–32]. However, the potential of DNA methylation as a predictor for anti-PD-1 therapy in GC has not yet been explored.

In this study, we collected tumor samples from advanced GC patients undergoing first-line anti-PD-1-based treatment and conducted a comprehensive DNA methylation analysis of over 850,000 CpG probes. By correlating these data with treatment outcomes, we aimed to evaluate the potential of DNA methylation as a biomarker for anti-PD-1 therapy. Additionally, we developed and validated a machine learning model based on differential methylation of response-related CpG probes to predict the therapeutic response to first-line anti-PD-1-based treatment in GC, with the goal of facilitating more precise treatment strategies.

Methods

Patient inclusion and sample acquisition

The workflow of this retrospective study is illustrated in Fig. 1. In the CpG feature discovery and model construction phase, we enrolled patients diagnosed with GC at the First Affiliated Hospital of Sun Yat-sen University between January 1, 2019, and January 1, 2024. In the model validation phase, GC patients diagnosed after February 1, 2024, were collected to serve as a temporally independent dataset for external validation. Formalin-fixed paraffin-embedded (FFPE) primary tumor tissues, collected before the initiation of treatment, were utilized for analysis. The inclusion criteria were as follows: (1) patients with histologically confirmed locally advanced GC or GC with distant metastases; (2) receipt of first-line treatment combining chemotherapy with immunotherapy, using anti-PD-1 antibodies (Nivolumab, Pembrolizumab, Camrelizumab, or Sintilimab); (3) availability of FFPE primary tumor tissues obtained through endoscopic biopsy or exploratory surgery prior to the start of treatment; (4) regular radiological assessments conducted following treatment. Exclusion criteria included the following: (1) HER-2 positive expression; (2) Inadequate quality of FFPE tissue for DNA methylation sequencing. 

Fig. 1.

Work flow of the study. This diagram illustrates the sequential steps taken in our study. In the CpG feature selection phase, methylation analysis of GC samples was performed using the Infinium MethylationEPIC BeadChip (850 K array) and Targeted Bisulfite Sequencing (TBS). Feature selection was conducted using SVM-RFE and LASSO. In the model construction phase, the final predictive model, iMETH, was developed using machine learning algorithms and evaluated for its ability to predict response to first-line anti-PD-1-based treatment. SHAP analysis was then employed to interpret the model. In the model validation phase, a temporally independent cohort of 28 samples underwent TBS for external validation

We collected clinical and pathological information, including age, gender, body mass index (BMI), smoking, alcohol consumption, tumor location, tumor differentiation grade, clinical TNM (cTNM) stage, treatment regimen, PD-L1 combined positive score (CPS), and MSI status. cTNM staging was based on the 8th edition of the American Joint Committee on Cancer (AJCC) staging system. Written informed consent for the collection and analysis of FFPE tissue was obtained from all participants. This study was approved by the Ethics Committee of the First Affiliated Hospital of Sun Yat-sen University (Approval No. [2022]715) and conducted in accordance with the biomedical research guidelines set forth in the Helsinki Declaration.

Assessment of treatment response and follow-up

Patients underwent baseline abdominal computed tomography (CT) scans prior to the initiation of treatment, with subsequent scans performed every three treatment cycles. Treatment response was assessed using the Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1) [33]. Patients were categorized as responders if they achieved a complete response (CR) or partial response (PR). Those with stable disease (SD), progressive disease (PD), or death due to early progression before the first assessment were classified as non-responders. Follow-up commenced at the start of treatment and included in-hospital or telephone evaluations during treatment and every three months thereafter. Progression-free survival (PFS) was defined as the time from the initiation of treatment to disease progression, while overall survival (OS) was defined as the time from treatment initiation to death. The final follow-up date was July 1, 2024.

Genome-wide DNA methylation profiling using Infinium MethylationEPIC BeadChip

Genome-wide methylation sequencing was performed to identify CpG probes associated with immunotherapeutic outcomes. Briefly, DNA was extracted from FFPE tissue samples using the DNeasy Blood & Tissue Kit (Qiagen, Germany). The purity and concentration of the extracted DNA were assessed with the Qubit 3.0 fluorometer. Subsequently, 500 ng of DNA from each sample was bisulfite-converted using the EZ DNA Methylation Kit (Zymo Research, USA). The converted DNA was then applied to Infinium MethylationEPIC BeadChip (850 K array) following the manufacturer’s guidelines and protocols provided by Sinotech Genomics Co., Ltd. (Illumina, USA). Data analysis was performed using the “ChAMP” package in R [34], where the methylation levels for each CpG probe were calculated from β values ranging from 0 (unmethylated) to 1 (fully methylated). Probes with detection p-values > 0.01, bead counts < 3 in more than 5% of samples, non-CpG probes, multi-hit probes, probes located on sex chromosomes (X and Y), and SNP-related probes were sequentially excluded, resulting in a final dataset of 726,710 CpG probes for further analysis. The “BMIQ” function was utilized to normalize the β value matrix, adjusting for type I and type II probe biases [35]. Singular Value Decomposition Analysis (SVA) was conducted to identify batch effects introduced by BeadChip Slides and Arrays [36], and the “ComBat” package in R was employed to correct for these batch effects. Finally, all CpG probes were annotated using the EPICanno.ilm10b4.hg19. Differentially methylated CpG probes (DMPs) were identified using champ.DMP function with criteria of |Δβ|≥ 0.10 and unadjusted p-value < 0.01. Genes associated with DMPs were subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses using the WebGestalt online tool (http://www.webgestalt.org) [37].

Targeted bisulfite sequencing and analysis

Targeted bisulfite sequencing (TBS) was utilized to evaluate the methylation levels of selected DMPs. Bisulfite sequencing primers were designed using MethPrimer. Genomic DNA (1 μg) was bisulfite-converted with the EZ DNA Methylation-Gold Kit (Zymo Research, USA). One-twentieth of the eluted product was employed as a template for polymerase chain reaction (PCR) amplification using the KAPA HiFi HotStart Uracil + ReadyMix PCR Kit (Kapa Biosystems, USA) for 35 cycles. The PCR products from multiple genes within each sample were pooled equally, 5′-phosphorylated, 3′-dA-tailed, and then ligated to barcoded adapters using T4 DNA ligase (NEB, USA). The barcoded libraries from all samples were sequenced on the Illumina platform. Initial quality control of the raw sequencing data was performed using FastQC (v0.11.7). Subsequent processing involved trimming low-quality reads, removing adapters, and filtering out short or unpaired reads using Trimmomatic (v0.36) [38]. Target sequences were aligned, and methylation levels at various sites on the amplicon were assessed using BSMAP v2.7.3 [39]. The methylation levels of the selected DMPs were then quantified.

Feature selection and construction of the prediction model

To reduce the dimensionality of the DMPs and identify the most informative features for predicting treatment response, we employed two machine learning algorithms: Support Vector Machine-Recursive Feature Elimination (SVM-RFE)[40] and Least Absolute Shrinkage and Selection Operator (LASSO) [41]. We first applied SVM-RFE to the DMPs identified from the 850 K array data. SVM-RFE is a wrapper-based feature selection technique that recursively eliminates features with the smallest contribution to the decision boundary of a support vector machine classifier. Specifically, it ranks features based on the absolute weights assigned by a linear SVM model, removes the least important ones, and repeats this process until the optimal feature subset is obtained. A fivefold cross-validation strategy was used to determine the number of features that achieved the best performance while avoiding overfitting. Subsequently, the selected features were further refined using LASSO regression, a penalized linear regression technique that performs both variable selection and regularization. LASSO imposes an L1 penalty on the regression coefficients, shrinking some of them to zero and thereby selecting a more parsimonious model. This two-step approach enabled the identification of a robust and compact set of CpGs for model construction.

After obtaining the methylation levels of the selected CpG probes, patients were divided into a training set and a testing set at a ratio of 7:3. We then constructed a CpG methylation prediction model for treatment outcomes using seven classical machine learning algorithms: Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (KNN), Adaptive Boosting (AdaBoost), Boosted Logistic Regression (LogitBoost), and Cancerclass. Logistic regression was used to construct a linear model as a control. The performance of each machine learning model was evaluated using Receiver Operating Characteristic (ROC) curves and confusion matrices. The optimal model, termed the iMETH model, was selected based on these evaluations. Additionally, PFS and OS were compared between responders and non-responders as predicted by the iMETH model.

Model interpretation

To interpret the results produced by the iMETH model, we utilized SHAP (SHapley Additive exPlanations) [42, 43], a game theory-based technique. Specifically, a SHAP explainer was created using the R package “DALEX.” SHAP values were used to measure the influence of each input feature on the predicted outcome. Higher SHAP values indicate that a CpG probe plays a more significant role as a predictor of treatment response. The results of the SHAP analysis were visualized using the “shapviz” package in R.

Temporally independent validation cohort

To evaluate the robustness and generalizability of the iMETH model, we retrospectively collected a temporally independent validation cohort of GC patients diagnosed after February 1, 2024. All patients were enrolled at the same center using identical inclusion criteria as the training and testing set. TBS was performed on FFPE tumor tissues to quantify the DNA methylation levels of the 20 CpG probes included in the iMETH model. The model’s ability to predict treatment response in this cohort was assessed using ROC curve analysis.

Statistical analyses

All statistical analyses were performed using R software (version 4.1.1). Continuous variables were compared using Student’s t-test or the Wilcoxon test, depending on data distribution, while categorical variables were evaluated using the chi-square test or Fisher’s exact test. For comparisons among multiple groups, the Kruskal–Wallis test was employed. The LASSO and SVM-RFE algorithms were implemented using the R packages “glmnet” and “e1071,” respectively. The “MIME” package was used to develop the predictive model for treatment response with machine learning algorithms.[44] Relevant parameters for the machine learning models are detailed in Additional file 1: Table S1. Survival analyses were conducted using the R packages “survival” and “survminer.” Kaplan–Meier curves were plotted, and the log-rank test was used to compare survival rates. As mentioned above, an unadjusted p-value of < 0.01 was used for the identification of DMPs. Statistical significance was set at p < 0.05 for other tests, and all tests were two-sided.

Results

Clinicopathological characteristics of patients

In CpG feature discovery and model construction phase, a total of 71 eligible patients were included, with a median follow-up time of 9.1 months. The cohort comprised 18 females (25.4%) and 53 males (74.6%), with a median age of 59 years (range, 31–88 years). Based on cTNM staging, the cohort included 3 patients with stage II, 26 with stage III, and 42 with stage IV GC, with the majority presenting with poorly differentiated tumors (63/71, 88.7%). The median number of anti-PD-1 therapy cycles was 6 (range, 2–18 cycles). The majority of patients (63/71, 88.7%) received the SOX regimen (S-1 + oxaliplatin) as the concurrent chemotherapy, while 5 patients received CapeOx (capecitabine + oxaliplatin), 1 patient with esophagogastric junction (EGJ) cancer received a paclitaxel + carboplatin regimen, 1 patient received mFOLFOX6, and 1 patient received S-1 + carboplatin. Of the 46 patients who underwent PD-L1 detection, 13 had a CPS of ≥ 5. MSI detection was performed for 51 patients, revealing 4 patients with MSI-H status (Table 1). Next, we compared the clinicopathological characteristics between responders and non-responders. Except for the fact that responders had a higher proportion of well-differentiated tumors (0% vs 17%, p = 0.045), there were no significant differences in other features (Additional file 1: Table S2).

Table 1.

Clinical information of patients in CpG feature discovery and model construction phase

Characteristic	N = 71
Age [M (range), year]	59 (31–88)
Gender [n (%)]
Male	53 (74.6)
Female	18 (25.4)
BMI [mean (SD), kg/m2]	22.9 (3.83)
Smoking [n (%)]
Yes	26 (36.6)
No	45 (63.4)
Alcohol consumption [n (%)]
Yes	14 (19.7)
No	57 (80.3)
Location [n (%)]
Upper 1/3	28 (39.4)
Middle 1/3	20 (28.2)
Lower 1/3	19 (26.8)
Whole	2 (2.8)
Remnant	2 (2.8)
Tumor invasion [n (%)]
T2	4 (5.6)
T3	10 (14.1)
T4a	47 (66.2)
T4b	10 (14.1)
Lymph node [n (%)]
N0	2 (2.8)
N1-3	69 (97.2)
Distant metastasis [n (%)]
M0	34 (47.9)
M1	37 (52.1)
cTNM stage [n (%)]
Ⅱ	3 (4.2)
Ⅲ	26 (36.6)
ⅣA	5 (7.0)
ⅣB	37 (52.1)
Differentiation [n (%)]
Well or Moderate	8 (11.3)
Poor	63 (88.7)
Anti-PD-1 agent [n (%)]
Camrelizumab	9 (12.7)
Pembrolizumab	2 (2.8)
Nivolumab	15 (21.1)
Sintilimab	45 (63.4)
Received cycles of anti-PD-1 therapy [M (range)]	6 (2–18)
Chemotherapy [n (%)]
SOX	63 (88.7)
CapeOx	5 (7.0)
Others	3 (4.2)
CPS [n (%)]
< 5	33 (46.5)
≥ 5	13 (18.3)
NA	25 (35.2)
MSI [n (%)]
MSI-H	4 (5.6)
MSS	47 (66.2)
NA	20 (28.2)
Follow–up time [M (range), months]	9.1 (1.7–26.8)
PFS [M (range), months]	6.1 (1.7–17.9)
OS [M (range), months]	Not reached (1.7–26.9)

M median, SOX S–1 + Oxaliplatin, CapeOx Capecitabine + Oxaliplatin, Others Paclitaxel + Carboplatin, mFOLFOX6 and S-1 + Carboplatin, CPS combined positive score, MSI microsatellite instability, PFS progression free survival, OS overall survival

According to RECIST 1.1, no patients achieved a CR at the first assessment; 45 patients (63.4%) achieved a PR, 10 patients (14.1%) had SD (2 patients achieved PR at the second assessment), and 16 patients (22.5%) experienced PD (Fig. 2a, b). Overall, 24 patients were classified as non-responders, while 47 patients were classified as responders. The median PFS from the start of treatment was 6.1 months, with a one-year survival rate of 79.4% (Fig. 2c, d).

Fig. 2.

Information on anti-PD-1-based treatment in the cohort. a Swimmer plot of 71 gastric cancer patients. b Radiological evaluation of 71 gastric cancer patients at the first assessment. c PFS from anti-PD-1 therapy start in months. d OS from anti-PD-1 therapy starting in months. CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease; PFS, progression-free survival; OS, overall survival

Identification of methylation CpG probes associated with anti-PD-1-based treatment response

To identify treatment-related methylation profiles across the whole genome, a cohort of 30 GC patients, comprising 19 responders and 11 non-responders, underwent Infinium MethylationEPIC BeadChip analysis. Compared to the unselected patients, this cohort is relatively younger in age (p = 0.01), but it covers a broader range of treatment regimens (p < 0.05). Other clinical characteristics remained consistent (p > 0.05, Additional file 1: Table S3). Following standard filtering procedures, data from 726,710 probes on the sequencing chips were retained for further analysis (Additional file 1: Figs. S1a, b, c). A total of 523 DMPs were identified (Δβ > 0.1, nominal p < 0.01), including 434 hypomethylated and 89 hypermethylated probes (Fig. 3a, Additional file 2: Table S4); however, none remained significant after Benjamini-Hochberg (BH) correction, which may be due to the limitation of sample size. Both hypomethylated and hypermethylated DMPs were uniformly distributed across all chromosomes, except the sex chromosomes (Additional file 1: Fig. S1d). These DMPs were mapped to various genomic regions: 34.1% of hypomethylated DMPs were located in gene bodies, 17.3% at transcription start sites (TSS), 9.4% in 5′ untranslated regions (5′UTR), 0.9% in 3′ untranslated regions (3′UTR), 1.3% in the first exon, and 0.2% at exon boundaries (Fig. 3b). Hypermethylated DMPs showed a similar distribution pattern, with slightly higher proportions in gene bodies (37.5%) and first exons (7.7%) (Fig. 3c). Regarding CpG island (CGI) locations, DMPs were predominantly found outside of CGIs, including shores (edges of CpG islands), shelves (regions between CGIs), and open sea (genomic regions not part of CGIs or their edges). Hypermethylated DMPs were more frequently located within CGIs (25.8% and 6.9%, respectively) (Fig. 3b, c).

Fig. 3.

Identification and enrichment analysis of differential methylation CpG probes. a The volcano plot illustrates differential methylation CpG probes between responders and non-responders to first-line anti-PD-1-based treatment. b and c Percentage of differential methylation CpG probes located in genomic regions and CpG island associated regions. d GO and KEGG enrichment analyses of 258 reference genes. CGI, CpG island; TSS, transcription start site; ExonBnd, exon boundary; IGR, intergenomic region; UTR, untranslated region; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; FDR, false discovery rate

Additionally, these DMPs were mapped to 258 genes. GO and KEGG enrichment analyses of these genes revealed three biological processes significantly enriched (FDR < 0.05): GO:0001667 (ameboidal-type cell migration), GO:0098984 (neuron to neuron synapse), and GO:0007163 (establishment or maintenance of cell polarity) (Fig. 3d).

Identification and selection of response-related DMPs

Among the 523 DMPs, we initially applied the SVM-RFE machine learning algorithm for feature selection. This process identified a combination of 51 DMPs with the highest prediction accuracy of 0.67 in fivefold cross-validation (Fig. 4a). These 51 DMPs were then refined using the LASSO algorithm, selecting those with the λ value that resulted in the minimum mean cross-validated error (Fig. 4b and Additional file 1: Fig. S2). Ultimately, we identified a subset of 20 DMPs most closely associated with anti-PD-1-based treatment response in GC (Table 2).

Fig. 4.

CpG feature selection. a SVM-RFE algorithm for feature selection of 523 DMPs. b Further screening of DMPs by LASSO algorithm. c Heatmap of the 20 DMPs’ methylation levels and clinical characteristics in 71 patients. DMP, differential methylation CpG probe; SVM, Support Vector Machine-Recursive Feature Elimination; LASSO, Least Absolute Shrinkage and Selector Operation

Table 2.

Information of 20 selected CpGs by SVM-RFE and LASSO algorithm

Probe ID	Chr	Address ID	RefGene	Gene ID	Location
cg13236889	chr5	92,630,114			IGR
cg21914317	chr20	82,619,496	SLC13A3	64,849	Body
cg01902489	chr8	16,732,164	ATP6V0D2	245,972	Body
cg23466166	chr8	5,679,567	PTK2	5747	5'UTR; Body
cg02192673	chr4	94,747,184	NPFFR2	10,886	1stExon; 5'UTR
cg23274420	chr1	27,632,113			IGR
cg26878154	chr1	81,674,408	BLACAT1	101,669,762	Body
cg01943225	chr9	48,641,913	MLANA	2315	Body
cg06692537	chr7	33,746,185	TMEM229A	730,130	TSS200
cg02992887	chr1	41,626,904	C1orf21	81,563	5'UTR
cg11368534	chr13	26,746,520	MYO16	23,026	5'UTR; 1stExon
cg02897989	chr1	54,687,167	PRRX1	5396	Body
cg00132255	chr1	54,713,894			IGR
cg07204874	chr1	79,651,256			IGR
cg24513480	chr3	33,641,117	SOX2OT	347,689	Body
cg08571680	chr11	66,791,922	TBX10	347,853	TSS1500
cg21226022	chr19	32,683,985	FCHO1	23,149	TSS200; 5'UTR
cg08094632	chr19	52,790,979			IGR
cg26639809	chr19	78,654,428			IGR
cg21370007	chr9	76,695,256	DAB2IP	153,090	Body

SVM-RFE Support Vector Machine-Recursive Feature Elimination, LASSO Least Absolute Shrinkage and Selector Operation, ExonBnd exon boundary, IGR intergenomic region, UTR untranslated region, TSS transcription start site

Subsequently, TBS was performed on the remaining patients to obtain methylation levels of these 20 DMPs. Principal Component Analysis (PCA) demonstrated no batch effects between the TBS data and the 850 K array data, indicating consistency between the two technologies and their suitability for combined analysis (Additional file 1: Fig. S3). Consequently, we acquired the methylation levels of the 20 DMPs from a total of 71 GC samples, as depicted in the heatmap in Fig. 4c. Based on treatment response, we divided the columns into four clusters to explore potential differences in methylation patterns across responders and non-responders. Notably, two of the four clusters consisted exclusively of responders, while non-responders were distributed across the remaining two clusters. These four clusters exhibited potentially distinct methylation patterns across the 20 DMPs. Except for the four MSI-H patients, who were exclusively found in responders, other clinicopathological features showed a relatively uniform distribution across the clusters. Overall, these 20 selected DMPs exhibit potential as predictive biomarkers for response to anti-PD-1-based treatment in GC.

Predictive model development

The 71 samples were randomly divided into a training set and a testing set at a ratio of 7:3. Seven machine learning algorithms were employed to construct prediction models. While four models (LogitBoost, AdaBoost, RF, and Logistic Regression) achieved perfect performance on the training set (AUC = 1.00) with confidence intervals of (1.00 ~ 1.00), indicating a potential risk of overfitting, their performance on the testing set was less impressive (AUC = 0.66 ~ 0.87), underscoring their limited generalizability. In contrast, the KNN model demonstrated superior performance, with AUC values of 0.99 (95% CI, 0.96 ~ 1.00) and 0.96 (95% CI, 0.85 ~ 1.00) on the training and testing sets, respectively. (Fig. 5a, b).

Fig. 5.

iMETH model construction and evaluation based on machine learning algorithms. a ROC curves of seven machine learning algorithms and logistic regression model in training set. b ROC curves of seven machine learning algorithms and logistic regression model in testing set. c and d PFS comparison between “responders” and “non-responders” predicted by iMETH model in training set and testing set. e and f OS comparison between “responders” and “non-responders” predicted by iMETH model in training set and testing set. ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; NB, naive bayes; SVM, support vector machine; RF, random forest; KNN, k-nearest neighbors; AdaBoost, adaptive boosting; LogitBoost, boosted logistic regression; PFS, progression-free survival; OS, overall survival; HR, hazard ratio

Survival analysis revealed that patients predicted as responders by the iMETH model experienced significantly longer PFS and OS in both the training and testing cohorts (all p < 0.05, HR = 0.11 ~ 0.23) (Fig. 5c, d, e, f). These findings were consistent with the actual response data (Additional file 1: Figs. S4a, b, c, d).

SHAP analysis for model interpretation

Next, we employed the SHAP method to interpret the iMETH model. Figure 6a and b presents the global feature importance analysis for the testing set using SHAP values. The analysis highlights that cg06692537, which is located at TSS200 of Transmembrane Protein 229A (TMEM229A), contributed most significantly to the iMETH model’s predictions. Specifically, lower methylation levels of cg06692537 were associated with a higher likelihood of benefiting from first-line anti-PD-1 therapy in patients with GC. For individual samples, the SHAP force plots provide a visual representation of how different methylation features influence the prediction of responders versus non-responders (Fig. 6c, d).

Fig. 6.

Interpretation of iMETH model by SHAP. a Beeswarm plot of CpG feature importance from SHAP analysis. b Bar plot of CpG feature importance from SHAP analysis. c SHAP waterfall plot of how the iMETH model arrived at its decision on an actual non-responder. d SHAP waterfall plot of how the iMETH model arrived at its decision on an actual responder. SHAP, SHapley Additive exPlanations

External validation of iMETH model in a temporally independent cohort

A total of 28 patients with advanced GC were enrolled in the temporally independent validation cohort. There were no statistically significant differences in clinicopathological characteristics (including age, sex, and tumor stage) between the validation cohort and the training or testing cohorts (all p > 0.05, Table 3). Among the 28 patients, 19 (67.9%) achieved PR at the first post-treatment CT evaluation, while 6 (21.4%) had PD and 3 (10.7%) had SD (Fig. 7a). There were no significant differences in clinicopathological characteristics between responders and non-responders (all p > 0.05, Additional file 1: Table S5). After obtaining the methylation levels of the 20 CpG probes from primary tumor tissues using TBS, the iMETH model was applied to predict treatment response. The ROC curve showed an AUC of 0.83 (95% CI, 0.74–0.96), demonstrating robust predictive performance (Fig. 7b).

Table 3.

Comparison of clinicopathological characteristics among training, testing and validation sets

Characteristic	Training set	Testing set	Validation set	p value
	n = 50	n = 21	n = 28
Age [mean (SD)]	59.2 (12.3)	56.6 (11.3)	59.0 (12.6)	0.70
Gender [n (%)]				0.67
Male	36 (72.0)	17 (81.0)	22 (78.6)
Female	14 (28.0)	4 (19.0)	6 (21.4)
BMI [mean (SD), kg/m2]	22.8 (3.9)	23.1 (3.7)	22.1 (2.2)	0.55
Smoking [n (%)]				0.30
Yes	18 (36.0)	8 (38.1)	15 (53.6)
No	32 (64.0)	13 (61.9)	13 (46.4)
Alcohol consumption [n (%)]				0.11
Yes	7 (14.0)	7 (33.3)	8 (28.6)
No	43 (86.0)	14 (66.7)	20 (71.4)
Location [n (%)]				0.36
Upper 1/3	19 (38.0)	9 (42.9)	6 (21.4)
Middle 1/3	13 (26.0)	6 (28.6)	9 (32.1)
Lower 1/3	16 (32.0)	4 (19.0)	13 (46.4)
Whole	1 (2.0)	1 (4.8)	0 (0.0)
Remnant	1 (2.0)	1 (4.8)	0 (0.0)
Tumor invasion [n (%)]				0.81
T2	3 (6.0)	1 (4.8)	0 (0.0)
T3	8 (16.0)	2 (9.5)	2 (7.1)
T4a	32 (64.0)	15 (71.4)	21 (75.0)
T4b	7 (14.0)	3 (14.3)	5 (17.9)
Lymph node [n (%)]				0.46
N0	1 (2.0)	1 (4.8)	0 (0.0)
N1-3	49 (98.0)	20 (95.2)	28 (100.0)
Distant metastasis [n (%)]				0.64
M0	25 (50.0)	9 (42.9)	11 (39.3)
M1	25 (50.0)	12 (57.1)	17 (60.7)
cTNM stage [n (%)]				0.83
Ⅱ	3 (6.0)	0 (0.0)	0 (0.0)
Ⅲ	19 (38.0)	7 (33.3)	10 (35.7)
ⅣA	3 (6.0)	2 (9.5)	1 (3.6)
ⅣB	25 (50.0)	12 (57.1)	17 (60.7)
Differentiation [n (%)]				0.91
Well or Moderate	6 (12.0)	2 (9.5)	2 (7.1)
Poor	44 (88.0)	19 (90.5)	26 (92.9)
Anti-PD-1 agent [n (%)]				0.14
Camrelizumab	6 (12.0)	3 (14.3)	0 (0.0)
Pembrolizumab	1 (2.0)	1 (4.8)	0 (0.0)
Nivolumab	11 (22.0)	4 (19.0)	3 (10.7)
Sintilimab	32 (64.0)	13 (61.9)	25 (89.3)
Chemotherapy [n (%)]				0.19
SOX	45 (90.0)	18 (85.7)	27 (96.4)
CapeOx	2 (4.0)	3 (14.3)	0 (0.0)
Others	3 (6.0)	0 (0.0)	1 (3.6)
CPS* [n (%)]				0.12
< 5	24 (77.4)	9 (60.0)	0 (0.0)
≥ 5	7 (22.6)	6 (40.0)	1 (100.0)
MSI* [n (%)]				0.65
MSI-H	3 (8.3)	1 (6.7)	0 (0.0)
MSS	33 (91.7)	14 (93.3)	18 (100.0)
Treatment response [n (%)]				0.99
Responder	33 (66.0)	14 (66.7)	19 (67.9)
Non-responder	17 (34.0)	7 (33.3)	9 (32.1)

SOX S-1 + Oxaliplatin, CapeOx Capecitabine + Oxaliplatin, Others Paclitaxel + Carboplatin, mFOLFOX6, S-1 + Carboplatin and Raltitrexed + Oxaliplatin, CPS combined positive score, MSI microsatellite instability

^*Not available for some patients

Fig. 7.

Validation of iMETH model in a temporally independent cohort. a Radiological evaluation of 28 gastric cancer patients at the first assessment. b ROC curve of iMETH model in validation set. CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease; ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval

Discussion

Immunotherapy has emerged as a promising treatment modality for GC, demonstrating more significant therapeutic effects compared to traditional chemotherapy and molecular targeted therapies. Its use has not only improved the pathological response rates but also yielded positive outcomes in overall prognosis [45, 46]. In our cohort of GC patients, the 1-year OS rate of first-line anti-PD-1 immunotherapy combined with chemotherapy was 79.4%, which was higher than the results of existing phase III clinical trials (approximately 55%) [4]. This higher rate may be attributed to the inclusion of resectable GC patients receiving perioperative treatment in our cohort, as well as the relatively short follow-up time.

In clinical practice, reliable predictive markers are essential to guiding precision therapy. However, in our study cohort, a considerable proportion of patients with CPS < 5 or MSS status still responded to anti-PD-1-based treatment. This indicates that these commonly used predictive biomarkers may not perform as well in the real world, particularly in terms of sensitivity.

To overcome these limitations, researchers worldwide are continually exploring novel biomarkers beyond conventional ones, including RNA sequencing biomarkers, immune microenvironment features, deep-learning-based pathology or imaging omics, and epigenetic or metabolomic markers [47–51]. Among these, DNA methylation, a well-established epigenetic modification, has shown potential as a predictive tool. Numerous studies have demonstrated that varying methylation profiles can differentiate therapeutic responses [29–31, 52]. Notably, this study is the first to associate the methylation profile with immunotherapy response in GC. We identified 523 DMPs, and GO and KEGG enrichment analyses reflected the potential biological functions. For instance, ameboidal-type cell migration is known to contribute to tumor cell invasiveness [53]; establishment of cell polarity is crucial for epithelial integrity, and its loss is associated with malignant transformation [54]; and although neuron-related synapse terms might appear unexpected, new evidence in breast cancer suggests that neuronal signaling pathways may influence cancer-immune interactions [55]. These findings provide a preliminary biological context for these DMPs, which warrant further investigation in the future.

Studies often involve hundreds or thousands of CpG methylation probes, which may constrain their clinical applicability. Complex classification models, such as those based on deep learning, may suffer from black-box issues, making their results less interpretable. Thus, balancing the number of predictive variables, the model’s discriminative power, and its interpretability is crucial. In our study, we addressed these challenges by employing machine learning algorithms to reduce the feature set to 20 through a rigorous stepwise screening process. Here we developed a TBS chip based on these 20 CpG methylation probes. The low cost and standardization ease of TBS enhance the model’s potential for clinical application. Furthermore, the iMETH model demonstrated robust fitting and generalization performance in both training and validation sets, likely due to the KNN algorithm’s non-parametric nature, which effectively captures subtle differences in methylation data (β values) among samples.

The SHAP algorithm enhanced the interpretability of the model by highlighting the specific contributions of the 20 selected DMPs to the decision-making process, providing valuable insights for future mechanistic studies. For instance, cg06692537, a key feature ranked first in the iMETH model, is located at the TSS200 region of the TMEM229A gene. It is widely acknowledged that high methylation in promoter regions often results in transcriptional repression or even gene silencing [56]. The SHAP analysis suggests that the low expression of TMEM229A may negatively impact the response to immunotherapy. TMEM229A, a member of the transmembrane transporter family [57], has been infrequently studied in cancer research. Notably, Zhang et al. demonstrated that TMEM229A can inhibit the progression of non-small cell lung cancer by suppressing the ERK pathway [58]. The ERK pathway is well-documented, and its inhibition has been associated with increased sensitivity of tumors such as colon cancer and melanoma to immunotherapy [59, 60]. Therefore, investigating whether TMEM229A influences GC immunotherapy response via the ERK pathway could be a potential future research direction. Other genes screened in the iMETH model have been associated with malignant tumor progression or prognosis in oncology, providing insights into potential therapeutic targets for GC [61–63]. Additionally, several CpG probes in the model are located at non-regulatory genomic regions. Increasing evidence suggests that CpG methylation in the Open Sea region can influence gene expression through remote regulatory mechanisms [56, 64], serving as potential novel regulatory elements that might indirectly or directly affect GC response to immunotherapy. Overall, there are currently few reports that explicitly link these genes or CpG probes with tumor immunotherapy, providing a new perspective for our future research.

Despite these promising findings, several limitations should be acknowledged. First, one major limitation of our study is that none of the DMPs identified remained significant after BH correction, largely due to the limited sample size and high dimensionality of the methylation array data. Although we adopted a relatively stringent threshold for DMP identification, the results should be interpreted with caution. Second, the study’s single-center and retrospective design may introduce selection bias, and the relatively small sample size and short follow-up duration, along with missing clinical and pathological data, may restrict the scope and accuracy of the analysis. Additionally, we only analyzed pre-treatment biopsy samples, excluding tissues collected during or after treatment. This limits our ability to assess how DNA methylation might reflect treatment efficacy over time. Therefore, future studies with larger sample sizes or integrated multi-omics data are warranted to validate these findings and improve statistical robustness.

Conclusions

In conclusion, our findings demonstrate that DNA methylation markers can serve as valuable predictors of first-line immunotherapy in GC. The iMETH model, which incorporates 20 DNA methylation CpG probes, effectively predicts patient responses to first-line anti-PD-1-based treatment. This model holds promise for guiding personalized treatment strategies and has the potential for practical application in clinical settings. 

Abbreviations

AdaBoost

Adaptive Boosting

AJCC

American Joint Committee on Cancer

AUC

Area under the curve

BMI

Body mass index

CGI

CpG island

CPS

Combined positive score

CR

Complete response

CT

Computed tomography

cTNM

Clinical TNM

DMP

Differential methylation probe

EBV

Epstein-Barr virus

FFPE

Formalin-fixed paraffin-embedded

GC

Gastric cancer

GO

Gene Ontology

HER-2

Human epidermal growth factor receptor 2

KEGG

Kyoto Encyclopedia of Genes and Genomes

KNN

K-nearest neighbors

LASSO

Least Absolute Shrinkage and Selector Operation

LogitBoost

Boosted Logistic Regression

MSI

Microsatellite instability

ORR

Objective response rate

OS

Overall survival

PCA

Principal Component Analysis

PCR

Polymerase chain reaction

PD

Progression disease

PD-1

Programmed death 1

PD-L1

Programmed death ligand 1

PFS

Progression-free survival

PR

Partial response

RECIST 1.1

Response Evaluation Criteria in Solid Tumors version 1.1

RF

Random Forest

SD

Stable disease

SHAP

SHapley Additive exPlanations

SVA

Singular Value Decomposition Analysis

SVM-RFE

Support Vector Machine-Recursive Feature Elimination

TBS

Targeted Bisulfite Sequencing

TMB

Tumor mutation burden

TMEM229A

Transmembrane Protein 229A

TSS

Transcription start sites

UTR

Untranslated regions

Authors’ contributions

Conceptualization, ZW2 and CC; methodology, ZW2; software, WT; validation, GL; investigation, WT; resources, ZW1 and GL; data curation, QL and ZZ; writing—original draft preparation, WT; writing—review and editing, ZW1 and ZW2; visualization, WT; supervision, SC; funding acquisition, ZW2 and ZZ. All authors read and approved the final manuscript. ZW1: Zhao Wang; ZW2: Zhixiong Wang.

Funding

This study is supported by the National Natural Science Foundation of China (Grant No. 81802342), Kelin New Star Project of the First Affiliated Hospital of Sun Yat‐sen University (Grant No. R08010), Guangzhou Municipal Science and Technology Project (Grant No. 2023A04J2208) and the Science and Technology Foundation of Panyu (Grant No. 2022-Z04-095).

Data availability

The sequencing data generated and analyzed during this study have been deposited in the NCBI Gene Expression Omnibus under accession numbers GSE305240 and GSE305511. Other data used and/or analyzed during the current study are available from the corresponding author, Zhixiong Wang, upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and reviewed and approved by the Ethics Committee for Clinical Research and Animal Trials of the First Affiliated Hospital of Sun Yat-sen University (Approval No. [2022]715). Written consent was obtained from all participants involved in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.