DNA methylation-related lncRNAs predict prognosis and immunotherapy response in gastric cancer

Abstract

Background

LncRNAs and DNA methylation are both key regulators of tumorigenesis and immune regulation. However, the interaction between lncRNA and DNA methylation, their regulation and their clinical and immune relevance in gastric cancer (GC) remain unclear.

Methods

In this study, we identified DNA methylation regulator-related lncRNAs through Pearson correlation analysis in The Cancer Genome Atlas datasets. Univariate Cox regression was used to screen DNA methylationrelated prognostic lncRNAs. Further, through least absolute shrinkage and selection operator Cox regression, a prognostic model based on 13 lncRNAs was established. Survival analysis and receiver operating characteristic curve analysis verified the accuracy of the model in predicting the survival of GC patients. Univariate and multivariate analyses also confirmed that the risk score obtained from the risk model could be applied as an independent prognostic factor for patients with GC. Furthermore, based on the risk score and other clinicopathological characteristics that can be used as independent prognostic factors, we constructed a nomogram that could accurately determine the survival time of each patient. In addition, a lncRNA score was constructed using a principal component analysis algorithm to quantify the DNA methylation-related lncRNA expression patterns of individual tumors.

Results

We found that a higher lncRNA score indicated a worse the prognosis and was associated with a reduced tumor mutation burden and immunosuppression. A low lncRNA score was related to an increase in neoantigen load and an increase in the anti-PDL1/CTLA4 immunotherapy response. Additionally, a low lncRNA score was related to a significant therapeutic advantage and clinical benefit.

Conclusions

This study describes a DNA methylation regulator-related lncRNA signature model, which provides a new approach for predicting therapeutic response and patient stratification in GC. Assessing lncRNA expression patterns in individual tumors will contribute to enhancing our understanding of tumor microenvironment infiltration and guide more effective immunotherapy strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00432-023-05234-8.

Introduction

Gastric cancer (GC) is the fifth most common cancer and the third most common cause of cancer death worldwide, with over 1 million estimated new cases and more than 700,000 deaths annually (Smyth et al. 2020; Joshi and Badgwell 2021). Due to its frequently advanced stage at diagnosis, the mortality rate of GC is high. In addition, because of population aging, there will be more cases of GC in future in some countries and regions (Smyth et al. 2020). Compared with traditional surgery, radiotherapy and chemotherapy, immunotherapy has become a powerful means of treating advanced gastric cancer. In addition, early screening and diagnosis are particularly important for patients with GC. Growing evidence shows that the discovery and application of molecular biomarkers can provide diagnostic and prognostic value (Zhang et al. 2020).

Genomes are extensively transcribed and give rise to thousands of long noncoding RNAs (lncRNAs), which are defined as RNAs longer than 200 nucleotides with no apparent coding capacity (Statello et al. 2021). LncRNAs have a variety of functions involved in tumor progression. In the nucleus, lncRNAs can regulate gene transcription, splicing, and nucleation, while in the cytoplasm, lncRNAs can act as miRNA sponges, interact with signaling proteins and modulate the translation of specific mRNAs (Engreitz et al. 2016). LncRNAs regulate gene expression at the transcriptional and posttranscriptional levels, and they also regulate epigenetics. Epigenetics involves DNA and RNA methylation, posttranslational histone modifications, and transcriptional regulation by lncRNAs (Loaeza-Loaeza et al. 2020). DNA methylation is a key layer of epigenetic regulation, and it is reportedly related to numerous biological processes, including the progression of GC (Parry et al. 2021). Methylation is dependent on the catalytic activity of DNA methyltransferases (DNMTs), and the active removal of DNMTs relies on the activity of ten–eleven translocation (TET) enzymes (Parry et al. 2021). Methylation regulation is one of the major mechanisms used to control lncRNA expression and tissue specificity (Wang et al. 2018). The promoter regions of lncRNAs often contain CpG islands, and DNA methylation can affect the methylation status of these regions, thereby affecting lncRNA transcription. For instance, He et al. reported that DNMT1-mediated lncRNA MEG3 methylation accelerates endothelial-mesenchymal transition in diabetic retinopathy through the PI3K/Akt/mTOR signaling pathway (He et al. 2021). DNA methylation also regulates lncRNA expression by affecting lncRNA stability. DNA methylation can recruit MBD proteins, which can bind to DNA methylation sites and further recruit RNA degrading enzymes to degrade lncRNA (Athanasopoulou et al. 2023). Similarly, lncRNA also regulates DNA methylation. LncRNA PVT1 recruits DNMT1 via EZH2 to the miR-18b-5p DNA promoter and suppresses the transcription of miR-18b-5p through DNA methylation in gallbladder cancer (Jin et al. 2020). LncRNA HOXA11-AS can scaffold the chromatin modification factors PRC2, LSD1, and DNMT1 to promote the proliferation and invasion of gastric cancer (Sun et al. 2016). The specific role of DNA methylation regulators in lncRNAs remains unclear. Therefore, understanding the mechanism of methylation-related lncRNAs in the development of GC may be useful for prognostic targets.

In this study, we investigated the expression patterns of 20 methylation regulators in gastric cancer. Methylation-related lncRNAs were identified through correlation calculations. Cox and least absolute shrinkage and selection operator analysis (LASSO) regression analyses were then performed to identify prognosis-related lncRNAs. We also established a prognostic risk model constructed from 13 lncRNAs. Finally, we further explored the relationship between these 13 lncRNAs and the tumor immune microenvironment, microsatellite instability and tumor mutation burden.

Materials and methods

Data acquisition and processing

RNA transcriptome data with the workflow type “HTseq-FPKM” and corresponding clinicopathological data in “xml” format were downloaded from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). GC patients with missing overall survival (OS) values were excluded. The clinical information extracted from the TCGA-STAD project is shown in Supplementary Table S1. Next, the “ComBat” algorithm of the “sva” package was used to correct batch effects from nonbiological technical biases. The mutation data were also acquired from TCGA in the annotated somatic mutation data type, and the workflow type was “VarScan2 Annotation.”

Selection of DNA methylation regulation genes and methylation-related lncRNAs

The lncRNA annotation file was acquired from the USCS website for annotation of the lncRNAs in the TCGA dataset to obtain an expression matrix (Rosenbloom et al. 2013). Based on previous studies, 3 erasers (TET3, TET2, TET1), 3 writers (DNMT3B, DNMT3A, DNMT1), and 14 readers (ZBTB33, ZBTB38, ZBTB4, MBD4, MBD3, MBD2, MBD1, NTHL1, MECP2, UNG, UHRF2, UHRF1, TDG, SMUG1) were included in the category of DNA methylation regulators. Methylation-related lncRNAs were screened by Pearson’s correlation analysis. The process used the criteria of | Pearson correlation coefficient |> 0.4 and P value < 0.001.

Assessing prognosis-related lncRNAs and constructing a prognostic model

The prognostic model construction method is described elsewhere (Guo et al. 2021). In brief, univariate Cox regression analysis was used to identify lncRNAs associated with methylation-related prognosis. We then utilized LASSO Cox regression analysis to identify the genes with the best prognostic value. The regression coefficients obtained in the regression model and lncRNA expression were used to calculate the risk score of patients with GC. The risk score calculation formula was as follows:

$$\text{Risk score}=\sum _{i=1}^n\text{Coefi}\times xi,$$

where coefi indicates the coefficients, and xi is the expression value of each lncRNA (Xu et al. 2021). The GC patients were stratified into low- and high-risk groups based on their risk score.

Unsupervised clustering of lncRNAs associated with DNA methylation-related prognosis

Unsupervised clustering methods were used to identify lncRNAs and gene expression patterns and for further analysis. The “ConsensusClusterPlus” (Wilkerson and Hayes 2010) R package was used to perform consensus clustering, and the calculation was repeated 1000 times to ensure cluster stability. The number and stability of clusters were determined by the consensus clustering algorithm.

Gene set variation analysis (GSVA) and single-sample GSEA (ssGSEA)

GSVA (Hänzelmann et al. 2013) enrichment analysis was performed to investigate the difference in biological processes between different lncRNA clusters by using the “GSVA” R package. GSVA is usually used to quantify the activity of biological pathways by an unsupervised and nonparametric method. The “c2.cp.kegg.v7.2.-symbols” file required for GSVA analysis was downloaded from the MSigDB database (http://www.gsea-msigdb.org/). Similarly, we used the ssGSEA algorithm in the GSVA R package to estimate the relative abundance of each immune cell in GC. In addition, the “clusterProfiler” R package was used to annotate lncRNA-related genes with a cutoff value of FDR < 0.05.

Differentially expressed genes (DEGs) between lncRNA clusters and functional enrichment analysis

The GC patients were divided into three categories based on 13 methylation-related lncRNAs, and an empirical Bayesian approach using the “limma” R package was used to determine DEGs between different clusters. The significance criterion of DEGs was an adjusted P value < 0.01. The “clusterProfiler” R package was used for Gene Ontology (GO) Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis of DEGs.

Generation of the methylation related-lncRNA score

To quantify the expression pattern of DNA methylation-related lncRNAs in an individual tumor, we constructed a scoring system to evaluate this expression pattern. First, a univariate Cox regression model was used to analyze the prognosis of each DEG. The extracted genes that were significantly related to prognosis were further analyzed. The expression of these genes was converted into a Z score. The DNA methylation-related lncRNA score was constructed by principal component analysis (PCA) and was termed the “lncRNA score”. PC1 and PC2 were selected as the signature scores. The method focuses on the set with the largest related or anti-correlated gene block in the set while downweighting contributions from genes that do not track with other set members (Zhang et al. 2020; Xu et al. 2021):

$$\text{LncRNA score}=\sum \text{PC}{1}_i+\text{PC}{2}_i,$$

where _i is the expression of 13 DNA methylation-related prognostic lncRNAs.

Tissue sample acquisition, RNA preparation and quantitative real-time PCR (qRT–PCR)

All GC tissues and adjacent normal tissue samples were obtained from the First Affiliated Hospital of Chongqing Medical University. This study was approved by the Ethics Committee of Chongqing Medical University and was in compliance with the relevant guidelines. Total RNA was obtained using TRIzol reagent (Takara, Japan) following the manufacturer’s protocol. Reverse transcription was performed with a PrimeScript RT Reagent Kit (Takara, #RR037A), and quantitative real-time PCR was carried out using TB Green Premix Ex Taq II (Takara, #RR820A). The 2^−△△CT method was used for data analyses. qRT–PCR amplification was performed on a Bio–Rad CFX96 machine (USA). Primer sequences for qRT–PCR were as follows:

CCNT2-AS1-F: 5ʹ-TGTTCCAGACCCCACACAGA-3ʹ;

CCNT2-AS1-R: 5ʹ- CCTGGTATCTGCTCACTCATCAA-3ʹ;

CYP1B1-AS1-F: 5ʹ-GGCAAAGGACCCCCATAGTC-3ʹ;

CYP1B1-AS1-R: 5ʹ-TCCCCTGGTAGAGCAGAGTC-3ʹ;

SENCR-F: 5ʹ- GTTACCTTGTCCACGCTCTC-3ʹ;

SENCR-R: 5ʹ-GTTTGAAGGTCGGTAGAGCC-3ʹ;

MSC-AS1-F: 5ʹ-TGAGGACCGCAGTACATACC-3ʹ;

MSC-AS1-R: 5ʹ-TACCAGTGACCACAATGGCT-3ʹ.

Statistical analyses

Paired or unpaired t tests were used for comparisons between the two groups. Kruskal–Wallis and one-way ANOVA tests were used to compare differences among three or more groups. Survival analyses were performed using the Kaplan–Meier method, and log-rank tests were used to identify the significance of differences. The univariate Cox regression model was used to calculate the hazard ratios (HRs) for different genes. Multivariate Cox regression models were used to determine independent prognostic factors. All statistical P values were two-sided, with P < 0.05 indicating statistical significance. All data processing was completed in R software version 4.1.0 (https://www.r-project.org/).

Results

Landscape of genetic variation of DNA methylation regulators in GC

Based on previous research findings regarding DNA methylation, including 3 writers, 3 erasers, and 14 readers, a total of 20 DNA methylation regulators were finally identified and analyzed in this study (Smith and Meissner 2013). The waterfall chart shows the mutations in DNA methylation regulators identified from the 433 samples in the TCGA-STAD project (Fig. 1A). The common mutation rate was not high; the three genes with the highest mutation rate were TET1 (5%), TET3 (5%) and ZBTB38 (4%), while UNG and TDG exhibited extremely low mutation rates in GC patients (0%). The investigation of the frequency of alterations in copy number variation (CNV) showed that the amplification frequency of MECP2, ZBTB38 and SMUG1 was high, while MBD1, MBD2 and UHRF1 had a widespread frequency of CNV deletion (Fig. 1B). The copy number variation of DNA methylation regulators in the TCGA database is provided in Supplementary Table S2. Next, we used univariate Cox regression to analyze the risk value of each DNA methylation regulator in GC. The comprehensive landscape of DNA methylation regulator interactions, risk significance for GC patients, and their regulator connection are depicted with the DNA methylation regulator network in Fig. 1C. Finally, we detected the differential expression of these 20 DNA methylation regulators in 375 GC samples and 32 adjacent normal tissues in the TCGA database. The expression of DNA methylation regulators in the TCGA database is provided in Supplementary Table S3. We found that except for ZBTB4 and MECP2, all others showed significantly higher expression in GC tissue compared with normal tissue, indicating that the disordered expression of methylation regulators may play an important role in the occurrence and progression of GC (Fig. 1D).

Fig. 1.

Landscape of genetic variation of DNA methylation regulators in gastric cancer. A The mutation frequency of 20 DNA methylation regulators in 433 patients with gastric cancer from TCGA-STAD cohort. Each column of the figure represents an individual patient. The upper bar plot represents TMB. The number on the right indicated the mutation frequency in each gene. The right bar plot showed the proportion of each mutation type. B The CNV frequency of DNA methylation regulators in TCGA STAD project. The loss frequency, blue dot; the gain frequency, red dot. C The interaction network among DNA methylation regulators. The size of the node indicates the strength of the interaction. D Boxplot shows the differential expression of 20 DNA methylation regulators between tumor and normal tissues in TCGA STAD cohort. Tumor, red; Normal, cyan. The upper and lower ends of the boxes represented interquartile range of values. The lines in the boxes represented median value, and black dots showed outliers. The asterisks represented the statistical P value. (*P < 0.05; **P < 0.01; ***P < 0.001)

Identification of DNA methylation-related lncRNAs in patients with GC

We annotated a total of 14,086 lncRNAs in the TCGA-STAD project. Through the Pearson correlation test and screening (| Pearson correlation coefficient |> 0.4 and P value < 0.001), a total of 451 DNA methylation regulator-related lncRNAs were obtained. The results are shown in Supplementary Table S4. DNA methylation regulators and their associated lncRNAs are shown in Fig. 2A. We then used univariate Cox regression analysis with a filter criterion of P < 0.05, and 30 prognostic lncRNA signatures were obtained (Fig. 2B). The results are shown in Supplementary Table S5. Furthermore, we visualized the expression correlation between DNA methylation regulators and prognostic lncRNAs (Fig. 2C). Most of these lncRNAs were positively correlated with ZBTB4 and MECP2, while they were negatively correlated with UHRF1, UNG, and TDG. Interestingly, all of these lncRNAs were differentially expressed between GC and adjacent tissue (Fig. 2D). The heatmap shows the expression of DNA methylation-related lncRNAs in GC and adjacent tissue (Fig. 2E).

Fig. 2.

Identification of DNA methylation-related lncRNAs in patients with GC. A Sankey diagram shows lncRNAs associated with 20 DNA methylation regulators. The left column represents DNA methylation regulators; the right column represents related lncRNA. B Forest plot of 30 candidate prognostic related lncRNAs. C Heatmap of the correlations between DNA methylation regulator and the 30 prognostic related lncRNAs. D, E Boxplot (D) and heatmap (E) show the differential expression of 30 prognostic related lncRNAs between tumor and normal tissues in TCGA STAD cohort. Tumor, red; Normal, cyan. *P < 0.05, **P < 0.01, and ***P < 0.001

Construction and validation of a risk model according to DNA methylation-related lncRNAs in GC patients

We obtained 30 DNA methylation-related prognostic lncRNAs through univariate Cox regression analysis. Based on this, we further performed LASSO regression analysis. The dashed perpendicular line illustrates the first-rank value of log (λ) with the minimum segment likelihood bias. We finally identified 13 DNA methylation-related lncRNAs that were significantly associated with overall survival (OS), and these were used to construct a prognostic model (Fig. 3A, B). The regression coefficient scores of the DNA methylation-related lncRNAs are shown in Fig. 3C. The forest plot shows that AC145285.6, AC007405.3, and AL355574.1 are protective factors with HRs < 1, while AL391152.1, AP001189.3, AC012409.3, SENCR, AL590705.3, MSC-AS1, AC129507.1, CYP1B1-AS1, AC022031.2, and CCNT2-AS1 are risk factors, with HRs > 1 in GC patients. According to the median risk score, patients were divided into low- and high-risk groups. The risk scores and groupings for each patient in the model are presented in Supplementary Table S6. We plotted survival curves by using the log-rank test and found that patients in the high-risk group had a worse prognosis than those in the low-risk group (P < 0.001) (Fig. 3E). In the high-risk group, the 1-year, 3-year, and 5-year survival rates were 0.692 (95% CI 0.622–0.768), 0.299 (95% CI 0.219–0.410), and 0.186 (95% CI 0.098–0.354), respectively. The survival rates were significantly higher in the low-risk group than in the high-risk group, at 0.836 (95% CI 0.782–0.895), 0.611 (95% CI 0.521–0.716), and 0.512 (0.383–0.686) for the 1-, 3-, and 5-year survival rates, respectively. Next, we constructed a ROC curve to test the prediction accuracy of the model. The AUC values at 1, 3, and 5 years were 0.681, 0.723 and 0.755, respectively, which indicated that the model had good accuracy (Fig. 3F). We then sorted the patients according to their risk score and analyzed the distribution of patients, with a median risk score as the cutoff value (Fig. 3G). As the risk score increased, more patients died, and the survival time was shorter (Fig. 3H). The heatmap showed that as the risk score increased, AL391152.1, AP001189.3, AC012409.3, SENCR, AL590705.3, MSC-AS1, AC129507.1, CYP1B1-AS1, AC022031.2, and CCNT2-AS1 expression increased; in contrast, AC145285.6, AC007405.3, and AL355574.1 expression decreased (Fig. 3I). In addition, we randomly divided the sample into testing set 1 and testing set 2 at a ratio of 1:1 to test the accuracy of the model (Figure S1). We found that in both testing sets, patients in the high-risk group fared worse than those in the low-risk group (Figure S1A, S1C). This is consistent with the previous results. The risk model had good AUC values in both groups (Figure S1B, S1D). In testing set 1 and testing set 2, after the patients were sorted by risk score (Figure S1E, S1F), as the risk score increased, the survival time shortened, and the number of deaths increased (Figure S1G, S1I). The lncRNA expression levels were also consistent with previous results (Figure S1I, S1J).

Fig. 3.

Construction and validation of a risk model according to DNA methylation-related lncRNAs in GC patients. A–C LASSO regression was performed, calculating the A, B minimum criteria and C coefficients. D 13 prognosis-related lncRNA were obtained by using LASSO regression and used for the construction of prognostic model. E Survival analysis of high-risk and low-risk groups. F ROC curves of 1, 2-, and 5-year survival prediction, with AUC = 0.681, 0.723, and 0.755, respectively. G, H The distribution of risk score and gene expression levels among patients in TCGA STAD cohort. I The expression of 13 prognostic lncRNA between high-risk and low-risk patients in TCGA training set

Independent prognostic value of the risk model

First, univariate and multivariate regression analyses were performed to confirm an independent prognostic factor. According to univariate Cox regression, age, stage, N stage, and risk score were significantly associated with OS (Fig. 4A). Multivariate Cox regression confirmed that age and risk score could act as independent prognostic indicators (Fig. 4B). Subsequently, a ROC curve was used to verify the accuracy of the risk score in evaluating prognosis. The AUC value of the risk score was 0.741, while the others did not exceed 0.65 (Fig. 4C). Decision curve analysis (DCA) also demonstrated that the risk model had elevated efficiency compared with other clinical characteristics (Fig. 4D). These results suggest that the risk score could be a good prognostic indicator for patients with GC. Subsequently, we constructed a more accurate nomogram for predicting patient survival based on clinicopathological characteristics and the risk score. A total score was calculated by adding scores corresponding to the various characteristics of the patient to assess the risk of death at 1, 3, and 5 years (Fig. 4E). For instance, patients with a total score of 440 had 1-year, 3-year, and 5-year mortality risks of 0.122, 0.349, and 0.414, respectively. The evaluated AUC values were 0.711, 0, 0.758 and 0.729 (Fig. 4F). Meanwhile, we observed from the calibration curve that the predicted survival time was quite similar to the actual survival time (Fig. 4G). In addition, we compared the survival of patients with different clinicopathologic characteristics between high- and low-risk groups. In nearly all cases, patients in the low-risk group had a better prognosis than those in the high-risk group (P < 0.001) (Fig. 5A–J), except for patients with stage I–II (Fig. 5K), T1–T2 (Fig. 5L), N0 (Fig. 5M), or M1 (Fig. 5N). The reason for these results may be due to the insufficient sample size, and the risk score may have a much better predictive effect for the prognosis of patients with advanced gastric cancer. We also used a heatmap to show the correlation between the risk score and clinicopathological characteristics of GC patients (Fig. 5O).

Fig. 4.

Independent prognostic value of the risk model. A Univariate and B multivariate analyses revealed that risk score was an independent prognostic predictor in the TCGA datasets. C ROC was performed to validate the superiority of the risk score and other clinicopathological features in predicting patient’ survival. D DCA was conducted to confirm the superiority of the risk score in clinical application. E Nomogram was plotted for the prediction of survival time in GC patients. F The ROC and G calibration curves were further plotted to determine the accuracy of the nomogram

Fig. 5.

Prognostic correlation between risk score and clinicopathological features. A–N Survival analysis was conducted in patients with various clinical characteristics in high and low-risk group. O Correlation between the risk model and patients’ clinical characteristic

DNA methylation prognosis-related lncRNA (DMP-lncRNA) expression patterns in the GC cohort

We first used 13 DMP-lncRNAs to establish a risk model, and we then further studied these 13 DMP-lncRNAs. The expression of these 13 DMP-lncRNAs in each GC patient is shown in Supplementary Table S7. First, we used unsupervised clustering to classify patients according to their expression patterns of the 13 DMP-lncRNAs. Considering that the maximum AUC increment of CDF was high within the group and low between the groups, the expression correlation of DMP-lncRNAs, k = 3, was determined as the number of clusters (Fig. 6A). We named them lncRNA cluster-A, lncRNA cluster-B and lncRNA cluster-C. Prognostic analysis of the three main expression subtypes revealed a particularly prominent survival advantage in the lncRNA cluster-B expression pattern, while lncRNA cluster-C had the worst prognosis (Fig. 6B). LncRNA cluster-B was characterized by the increased expression of AL355574.1, AC007405.3, AC02201.2 and AL391152.1 and presented variable decreases in other clusters. LncRNA cluster-C showed high expression of AP001189.3, AC012409.3, SENCR, MSC-AS1, AC129507.1 and CYP1B1-AS1. LncRNA cluster-A was somewhat in between, and there was no obvious high or low gene expression pattern (Fig. 6C). There appeared to be no clear correlation between the expression of these genes and clinical features. We next performed GSVA enrichment analysis to explore differences in the biological behaviors among the three DMP-lncRNA expression patterns. LncRNA cluster-A was markedly enriched in genes related to apoptosis, nod-like receptor signaling pathways, natural killer cell-mediated pathways, cytokine–cytokine receptor interactions and JAK/STAT signaling pathways (Fig. 6D). The cluster with the best prognosis, lncRNA cluster-B, presented enrichment associated with homologous recombination, spliceosome and base excision repair normal biological process pathways (Fig. 6E). The lncRNA cluster-C enrichment pathway was significantly related to cancer and was mainly enriched in TGF-β, WNT, MAPK, regulation of the actin cytoskeleton, pathways in cancer, and other related pathways (Fig. 6F). The top 20 enriched pathways are shown in Supplementary Table S8.

Fig. 6.

DNA methylation prognosis-related lncRNAs (DMP -lncRNAs) expression pattern in GC cohort. A Consensus matrixes of all GC cohorts for k = 3 show the stability of clustering through 1000 hierarchical clustering. B Survival analysis of three DNA methylation regulator-related lncRNA cluster, including 241 cases of cluster-A, 76 cases cluster-B, and 54 cases of cluster-C. Log-rank test, P = 0.029. C Unsupervised clustering of 13 lncRNA in the GC cohort. The lncRNA cluster and clinicopathological features of the patients were used as patient annotations. Red represented high expression of lncRNA and cyan represented low expression. D–F The GSVA method was used to quantify the activity of biological pathways between D lncRNA cluster-A and lncRNA cluster-B, and between E lncRNA cluster-A and lncRNA cluster-C and between F lncRNA cluster-B and lncRNA cluster-C

Construction and analysis of lncRNA score and its relationship with tumor mutation burden (TMB)

Although the consensus clustering algorithm based on DMP-lncRNA expression classified patients into three expression patterns, the underlying expression perturbations and genetic alterations in these patterns were unclear. We used the PCA method to construct a DNA methylation prognosis-related lncRNA score based on DMP-lncRNAs, referred to as the lncRNA score. The lncRNA score of each patient is presented in Supplementary Table S9. Based on the optimal cutoff value of 1.75, determined by the “survminer” package, patients were classified into low- or high-lncRNA score groups. The heatmap shows the correlation between the expression of lncRNAs in patients with high and low lncRNA scores and their clinical status, as well as the correlation with the lncRNA cluster (Fig. 7A). The previous analysis showed that lncRNA cluster-C had a poor prognosis. Here, lncRNA cluster-C was mainly concentrated in the high lncRNA score group. Next, we aimed to further determine whether the lncRNA score predicted patient outcomes. The results showed that patients with high lncRNA scores had significantly worse survival (P = 0.01) (Fig. 7B). Correlation analysis with the lncRNA clusters revealed that in cluster-C, the lncRNA score had the highest median, while in cluster-B, it was the lowest; in cluster-A, the lncRNA score was the most widely distributed, with a value between the two (Fig. 7C). This result was consistent with the previous findings (Fig. 6B) showing that higher lncRNA scores were associated with poorer outcomes. We further analyzed the correlation between the lncRNA score and TMB. We found that in the low lncRNA score group, the TMB was higher, while in the high lncRNA score group, the TMB value was lower (P = 2e−10) (Fig. 7D). The TMB value of each GC patient in the TCGA database STAD project is shown in Supplementary Table S10. Similarly, the correlation analysis also found that lncRNA score was negatively correlated with TMB (P = 22e−16, R = − 0.44) (Fig. 7E). We divided the patients with GC into high-TMB (H-TMB) and low-TMB (L-TMB) groups according to their TMB value and found that patients with high TMB had a better prognosis and patients with low TMB had a worse prognosis (Fig. 7F). By combining the TMB and lncRNA score analyses, we found that in both the H-TMB and L-TMB groups, those with low lncRNA scores had a better prognosis. Patients with H-TMB and a low lncRNA score had the best prognosis, while patients with L-TMB and a high lncRNA score had the worst prognosis (Fig. 7G). We then used the “Maftools” package to analyze differences in the distribution of somatic mutations between those with low and high lncRNA scores in the TCGA-STAD cohort. The TMB of the low lncRNA score group (Fig. 7H) was more extensive than that of the high lncRNA score group (Fig. 7I), with the rate of the 10th most significantly mutated gene being 22% versus 13%. Quantitative TMB analysis also confirmed the previous result that patients whose tumors had a low lncRNA score were significantly associated with higher TMB.

Fig. 7.

Construction and analysis of lncRNA score and its relationship with TMB. A The lncRNA score, lncRNA cluster and clinicopathological features of the patients were used as patient annotations. B Survival analysis of lncRNA score between high and low groups. Log-rank test, P = 0.01. C Differences in lncRNA score among three lncRNA clusters in GC patients. The Kruskal–Wallis test was used to compare the statistical difference between three gene clusters (P < 0.001). D Comparison of TMB values between high and low lncRNA score groups (P < 0.001). E Correlation test between lncRNA score and TMB value (R = − 0.44, P < 0.001). F Survival analysis of high-TMB and low-TMB groups. Log-rank test, P < 0.001. G Survival analysis of lncRNA score combined with TMB value

Predictive value of lncRNA scores in the tumor immune microenvironment and immunotherapy

Although immunotherapy has shown improved survival in the treatment of advanced gastric cancer (Janjigian et al. 2021), there is still an urgent need to determine which type of gastric cancer patient will benefit the most. Therefore, we further evaluated the response of lncRNAs to immunotherapy. We first estimated the immune cell content of GC samples using CIBERSORT algorithms. The results are shown in Supplementary Table S11. Differences in immune cell content between the high- and low-score groups were then analyzed. Regulatory T cells (Tregs), gamma delta T cells, resting NK cells, activated NK cells, M0 macrophages, M2 macrophages, and resting dendritic cells were included in this analysis. We found that gamma delta T cells, resting NK cells, activated NK cells, and M0 macrophages were significantly higher in the low lncRNA score group than in the high lncRNA score group. M2 macrophages and resting dendritic cells were higher in the high-score group (Fig. 8A). This may suggest that patients with lower scores are more responsive to immunotherapy. The correlations between lncRNA score and immune cells are presented in Fig. 8B. Immune checkpoint inhibitor therapy, represented by CTLA-4/PD-1, is undoubtedly a major breakthrough in antitumor therapy in recent years (Chong et al. 2021). To verify whether the lncRNA score can be used as an indicator for evaluating immunotherapy, we obtained immunotherapy score data from TCIA (https://tcia.at/) and compared the immunotherapy efficacy between the two groups. The results are shown in Supplementary Table S12. The results showed that immunotherapy was more efficacious in CTLA4-negative, PD-1-negative patients and CTLA4-positive, PD-1-negative patients with low lncRNA scores. However, in the CTLA4-positive, PD-1-positive group, a higher lncRNA score was associated with a higher immunophenoscore (IPS) (Fig. 8C–F). These findings indirectly demonstrated that the lncRNA score may play a crucial role in mediating the immune response. In addition, we found that PD-L1 expression was significantly higher in the high lncRNA group than in the low lncRNA group (P = 9.3e−15) (Fig. 8G). We also analyzed the relationship between lncRNA score and microsatellite instability. A significantly lower lncRNA score was found in samples of the MSI-L subtype than in those of the other two subtypes (Fig. 8H, I). There were significantly more deaths of patients with a high-score lncRNA array than of those with a low-score lncRNA array. Based on the above (Fig. 8J, K), there was significantly higher mortality in the high-score group than in the low-score group, which may suggest that the reason for the poor prognosis of high lncRNA scores is due to tumor cell immune escape. Finally, we selected the annotated lncRNAs in the TCGA and GEO databases to verify the expression of a small sample of GC tissues and adjacent tissues. The annotated lncRNAs included CCNT2-AS1, CYP1B1-AS1, SENCR and MSC-AS1. CCNT2-AS1 (P = 0.407, n = 16) (Figure S2A, S2B), CYP1B1-AS1 (P = 0.118, n = 16) (Figure S2C, S2D) and SENCR (P = 0.085, n = 15) (Figure S2E, S2F) were slightly higher in the adjacent tissues than in the tumor tissues, while MSC-AS1 expression was lower in the adjacent tissues (P = 0.822, n = 16) (Figure S2G, S2H). The lack of significant differences may have been due to an insufficient sample size. This result is basically consistent with the results of the combined analysis of the tumor data from the TCGA database and GTEx normal samples. The expression levels of CCNT2-AS1 (Figure S2I), CYP1B1-AS1 (Figure S2J), SENCR (Figure S2K) and MSC-AS1 (Figure S2L) in TCGA tumor samples and GTEx normal samples are shown in Figure S2I-2L. Survival analysis was performed for these lncRNAs using data from the TCGA and GEO databases. In the TCGA dataset, CCNT2 (HR = 1.9) (Figure S2M), CYP1B1-AS1 (HR = 1.6) (Figure S2N), SENCR (HR = 1.6) (Figure S2O) and MSC-AS1 (HR = 2) (Figure S2P) were found to be risk factors for GC, and the higher the expression, the worse the prognosis. The results from the GEO database including the GSE14210, GSE15459, GSE22377, GSE29272, GSE51105 and GSE52254 datasets (a total of 631 samples) were consistent with these findings. A higher expression of CCNT2-AS1 (HR = 1.1, P = 0.39) (Figure S2Q), CYP1B1-AS1 (HR = 1.6, P = 0.016) (Figure S2R), SENCR (HR = 1.6, P = 0.00017) (Figure S2S) and MSC-AS1 (HR = 1.6, P = 4.8e−10) (Figure S2T) indicated a worse prognosis.

Fig. 8.

Predictive value of lncRNA score in tumor immune microenvironment and immunotherapy. A Difference analysis of tumor immune infiltrating cells between high lncRNA score and low groups. B Correlation analysis of lncRNA score and immune cell infiltration. Negative correlation, cyan; Positive correlation, red. C–F The violin diagram showed the relationship between immunotherapy response and lncRNA score group in patients with in the C CTLA4-negative/PD1-negative group, D CTLA4-negative/PD1-positive group, E CTLA4-positive/PD1-positive group, and F CTLA4-negative/PD1-negative group, respectively. G Differential expression of PD-L1 in high and low lncRNA scores between the two groups. H The histogram and I boxplot, respectively, showed the survival status of patients with high and low lncRNA scores group. J Rate of microsatellite instability state in high or low lncRNA score groups in the TCGA-STAD cohort. K Difference analysis of microsatellite instability in lncRNA score high and low groups

Discussion

In recent years, an increasing number of studies have shown that DNA methylation regulators play an important role in antitumor immunomodulatory and immunotherapeutic responses. Blocking DNA methylation with the demethylating agent decitabine, which has been approved by the FDA, was shown to enhance immune checkpoint blockade (ICB) with anti-PD-L1 to rejuvenate CD8⁺ T cells and control tumor growth (Ghoneim et al. 2017). It has been confirmed that in vitro and in vivo demethylation drug therapy can restore the immunogenicity of cancer cells, and the cytotoxic activity of CD8⁺ T cells can enhance their infiltration into tumor tissues (Chiappinelli et al. 2015; Roulois et al. 2015). While lncRNAs are regulated by methylation, they also have a regulatory effect on tumor immunity and the tumor microenvironment. Since most studies have focused on a single regulatory factor or a single TME cell type, the overall characteristics of TME infiltration mediated by the combined effects of multiple DNA methylation-related lncRNAs have not been fully recognized (Van Acker et al. 2021). Identifying the role of different DNA methylation regulator-related lncRNA expression patterns in predicting patient prognosis and immune cell infiltration will help enhance our understanding of TME infiltration characterization and improve the effectiveness of immunotherapy strategies (Bockhorst et al. 2021).

In this study, we explored the mutation and expression of 20 methylation regulatory factors in GC. All investigated DNA methylation regulators except ZBTB4 and MECP2 were significantly overexpressed in GC. This suggests that the disordered expression of methylation regulators may be one of the causes of tumorigenesis. Accumulating studies have shown that DNA methylation modifications and lncRNAs play an indispensable role in innate immunity and tumorigenesis (Wang et al. 2018; Ponnusamy et al. 2019; Yu et al. 2021). However, the clinical relevance of the DNA methylation modification of lncRNAs in GC remains unclear. Here, through Pearson correlation analysis, we identified 451 lncRNAs related to DNA methylation. Furthermore, through univariate Cox regression and LASSO regression analysis, 13 DMP-related lncRNA signatures were identified and used to construct a risk model. Surprisingly, we found that some lncRNAs, such as CYP1BI-AS1, SENCR, and CCNT2-AS1, were downregulated in tumors compared with normal tissue. However, in our risk model, these are high-risk genes. We hypothesize that these lncRNAs do not exert their activity in normal tissue. In tumors, upstream DNA methylation regulators are activated, resulting in the transcriptional inhibition of lncRNAs and thus decreased expression. The high expression of some lncRNAs was not shown to induce immune activation; thus, these were identified as high-risk genes. Alternatively, genetic mutations might cause genes to exhibit different characteristics in tumors and normal tissue. Zhang and Zha et al. reported similar findings (Cao et al. 2021; Zha et al. 2021). Further research is needed to explore the possible mechanisms. In short, our model can accurately stratify patients with GC, and patients can be divided into high-risk and low-risk groups according to the risk score. Patients in the low-risk group had a significantly better prognosis than those in the high-risk group. The accuracy of the model was verified by survival curves and ROC curves. In addition, we used the risk score in conjunction with patient clinicopathological features to construct a nomogram to better predict patient outcomes.

We also conducted unsupervised cluster analysis based on the expression of the 13 DMP-lncRNAs and identified three expression patterns with different characteristics. Cluster-C was mainly enriched in cancer-related pathways such as the TGF-β, WNT, and MAPK signaling pathways. Therefore, it was not surprising that cluster-C had the worst prognosis among the three clusters. Cluster-C was mainly characterized by the high expression of CYP1B1-AS1, AP001189.3, AC012409.3, SENCR, MSC-AS1 and AC129507.1, which indicates that these genes may be used as tumor biomarkers for poor prognosis in GC patients. Among them, and in line with our study, the ferroptosis-related lncRNA AP001189.3 is used as a prognostic marker in colon cancer (Cai et al. 2021). CYP1B1-AS1 positively correlates with TGFβ1 in triple-negative breast cancer patients (Vishnubalaji and Alajez 2021). LncRNA SENCR reportedly promotes the progression of non-small-cell lung cancer through sponging miR-1-3p (Cheng et al. 2021). MSC-AS1 has been shown to play a role in promoting cancer progression and is a tumor biomarker (Yao et al. 2020; Liu et al. 2021). Cluster-A was mainly enriched in apoptosis, nod-like receptors, natural killer cell-mediated cytotoxicity, toll-like receptor signaling, cytokine–cytokine receptor interactions, and chemokine signaling pathways, hinting at the underlying regulatory mechanisms of the DNA methylation of lncRNAs in the progression of GC through immune pathways.

Furthermore, considering the individual heterogeneity of DMP-lncRNA expression patterns, it was necessary to quantify the DMP-lncRNA expression profiles of individual tumors. Therefore, we constructed a DMP-lncRNA scoring system based on the PCA algorithm, and this was called the lncRNA score. According to the lncRNA score, patients could be divided into two groups—those with high and low scores. The higher the score, the worse the prognosis. Interestingly, patients in the high-score cohort had lower TMB values, whereas in tumors, higher TMB was associated with a better prognosis. This may suggest that patients with a lower lncRNA score have a higher TMB, and their tumors would more likely be recognized by the immune system, thereby activating the immune system and leading to a better prognosis. Subsequent studies have confirmed a higher number of mutated genes and a greater abundance of immune cell infiltrates in patients with lower lncRNA scores. We also found that the combined analysis of TMB and lncRNA score could better predict patient prognosis. More importantly, PD-L1 was highly expressed in patients with high lncRNA scores, which may suggest that patients with high scores are more prone to experience immune escape. We also investigated the use of lncRNAs to assess the response to CTLA4 and PD-1 inhibitory treatments. It is helpful to indicate the sensitivity of patients with different scores so that different immune checkpoint inhibitors can be applied. Our findings provide new insights into the development of novel therapeutic strategies.

In conclusion, we constructed a risk model using 13 DMP-lncRNAs and a lncRNA scoring system, which are valuable tools for predicting patient survival and immunotherapy effects and may help to promote personalized GC immunotherapy in future.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

XG designed the study. XG and YW collected and analyzed data. LZ and HL collected tissues and conducted experiments. XG and YW wrote and revised the manuscript. KQ was responsible for supervising the study. All authors read and gave final approval of the manuscript.

Funding

This work was supported by the academic leader Reserve Talent Fund project of the First Affiliated Hospital of Chongqing Medical University (03010203XKTS085).

Data availability

Transcriptomic, matched clinical data and MAF files were downloaded from the TCGA database (https://portal.gdc.cancer.gov/).

Declarations

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.