Skip to main content
NCBI home page
Annals of Medicine logoLink to Annals of Medicine
. 2024 Nov 11;56(1):2426758. doi: 10.1080/07853890.2024.2426758

Integrating transcriptomic data and digital pathology for NRG-based prediction of prognosis and therapy response in gastric cancer

Qiuyan Sun a,b,c,*, Tan Li d,*, Zheng Wei a,b,c, Zhiyi Ye a,b,c, Xu Zhao e,✉, Jingjing Jing a,b,c,✉
PMCID: PMC11556273  PMID: 39527470

Abstract

Background

Cancer is characterized by its ability to resist cell death, and emerging evidence suggests a potential correlation between non-apoptotic regulated cell death (RCD), tumor progression, and therapy response. However, the prognostic significance of non-apoptotic RCD-related genes (NRGs) and their relationships with immune response in gastric cancer (GC) remain unclear.

Methods

In this study, RNA-seq gene expression and clinical information of GC patients were acquired from The Cancer Genome Atlas and the Gene Expression Omnibus databases. Cox and LASSO regression analyses were used to construct the NRG signature. Moreover, we developed a deep learning model based on ResNet50 to predict the NRG signature from digital pathology slides. The expression of signature hub genes was validated using real-time quantitative PCR and single-cell RNA sequencing data.

Results

We identified 13 NRGs as signature genes for predicting the prognosis of patients with GC. The high-risk group, characterized by higher NRG scores, demonstrated a shorter overall survival rate, increased immunosuppressive cell infiltration, and immune dysfunction. Moreover, associations were observed between the NRG signature and chemotherapeutic drug responsiveness, as well as immunotherapy effectiveness in GC patients. Furthermore, the deep learning model effectively stratified GC patients based on the NRG signature by leveraging morphological variances, showing promising results for the classification of GC patients. Validation experiments demonstrated that the expression level of SERPINE1 was significantly upregulated in GC, while the expression levels of GPX3 and APOD were significantly downregulated.

Conclusion

The NRG signature and its deep learning model have significant clinical implications, highlighting the importance of individualized treatment strategies based on GC subtyping. These findings provide valuable insights for guiding clinical decision-making and treatment approaches for GC.

Keywords: Gastric cancer, non-apoptotic regulated cell death, deep learning, prognosis, therapy

1. Introduction

Gastric cancer (GC) is a common malignant neoplasm that poses a significant threat to human health. In recent years, the introduction of surgery, chemotherapy, and immunotherapy has led to a significant improvement in the overall survival (OS) rate of GC [1]. However, due to late diagnosis, drug resistance, and immune evasion [2], GC continues to rank as the fifth most frequently diagnosed cancer with the fourth highest mortality rate worldwide [3]. Therefore, it is urgent to develop a robust classifier that can accurately identify patients who will benefit the most from precision therapy, thereby improving the survival rate of patients with GC.

Cell death is categorized into two main types: accidental and regulated cell death (RCD). Based on distinct morphological, biochemical, immunological, and genetic features, RCD is further divided into apoptotic and non-apoptotic subtypes, with the latter encompassing autophagy, ferroptosis, pyroptosis, and necroptosis [4,5]. In addition to regular apoptosis, the relationship between non-apoptotic RCD and cancer progression as well as the response to therapy has attracted the attention of researchers. The heightened susceptibility of tumors to ferroptosis, as seen in pancreatic cancer, has demonstrated substantial inhibition of both tumor development and progression [6]. Furthermore, ferroptosis can augment metabolic and inflammatory modulation of tumor-associated macrophages, leading to enhanced tumor-killing activity [7]. Targeting PIKfyve to inhibit autophagy has been found to enhance responsiveness to immune checkpoint blockade in prostate cancer [8]. Additionally, inducing necroptosis through the inhibition of Aurora Kinase A also showed promise in slowing in situ tumor growth in mice with pancreatic cancer [9]. Diverging from pan-cell death or specific forms of cell death, focusing on non-apoptotic RCD-related genes (NRGs) potentially opens up new avenues for innovative cancer therapies and serves as a complementary extension to existing research efforts.

Deep learning is a robust and time-efficient method for analyzing digital whole-slide images (WSIs), making it a valuable tool for advancing digital pathology [10,11]. The availability of scanned copies of WSIs provides a solid foundation for the application of deep learning techniques in pathological image analysis and has demonstrated excellent performance [10]. Tumor morphology reflects genetic changes in tumor cells, and the valuable information it provides can be used to predict tumor biology, clinical behavior, and treatment response [12,13]. Many researchers have focused on identifying computer-assisted image features extracted from routine hematoxylin and eosin (H&E)-stained images to aid disease prognosis prediction and facilitate clinical decision-making [14,15]. Deep learning can extract high-dimensional data from standard medical images, enabling its utilization in diverse clinical applications, including predicting microsatellite instability status [16], immune subtypes [17], and prognosis [14] in various tumors, which provides references for our study.

In this study, we systematically investigated the multilevel relationships between NRGs and GC. A prognostic signature consisting of 13 NRGs was constructed to classify patients with GC, which exhibited significant correlations with prognosis, immune microenvironment status, and sensitivity to chemotherapy/immunotherapy. Moreover, we developed a deep learning model based on digital pathology to predict the NRG signature, enabling its practical application in clinical settings.

2. Materials and methods

2.1. Data and sample collection

Gene expression profiles and corresponding clinical information on GC were obtained by downloading data from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). The TCGA cohort comprised 375 tumor specimens (STAD) and 32 normal specimens. The GSE84437 dataset extracted from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) was used as the test dataset to validate the predictive efficiency of the NRG signature. Single-cell RNA sequencing (scRNA-seq) data were retrieved from the GEO database (GSE183904). A total of 371 NRGs were obtained from the PathCards (https://pathcards.genecards.org/) and Harmonizome (https://maayanlab.cloud/Harmonizome/) databases (Supplementary Table 1).

Twenty-six pairs of GC tissue and adjacent normal tissue samples were collected from patients who underwent surgery at the First Hospital of China Medical University. This study was approved by the Ethics Review Committee of the First Hospital of China Medical University, and written informed consent was obtained from each participant.

2.2. Cluster analysis

An unsupervised clustering method called ‘Pam’ was applied to identify distinct clusters based on NRG expression. This procedure utilized the ‘ConsensusClusterPlus’ package and was repeated 1000 times to ensure the stability of the classification process. In the TCGA-STAD cohort, the optimal number of clusters was identified using the cumulative distribution function curve, consensus score, and consensus matrix. To evaluate prognosis, Kaplan–Meier (K–M) survival analysis was conducted to compare survival outcomes between the two clusters. Additionally, a heatmap was generated to visualize the correlations between clusters and various clinicopathological features, such as histological grade, pathological stage, T stage, N stage, M stage, age, and gender.

2.3. Analysis of differentially expressed genes (DEGs)

DEGs were identified from distinct clusters and subgroups using the ‘limma’ package with the criteria of |log fold change (FC)| > 1 and false discovery rate (FDR) < 0.05. Subsequently, the identified DEGs were subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis using the ‘ggplot2’ package.

2.4. Development, assessment, and validation of the NRG signature

Cox regression analysis was used to evaluate the prognostic value of DEGs. To construct an NRG risk score model (NRG signature), Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was performed and 13 DEGs were identified. The risk score for each patient was determined by calculating the expression levels of DEGs multiplied by their respective coefficients, based on the following formula: risk score =∑ (exp (Xi) × coef (Xi)), where ‘exp (Xi)’, ‘coef (Xi)’, and ‘i’ represented the expression level, the coefficient, and the number of signature genes, respectively. The patients from the TCGA-STAD cohort were assigned to the training group (n = 371), while the GSE84437 patients were assigned to the testing group (n = 433), and the signature was constructed in the training group. Using the median risk score as a threshold, the patients in the training and testing group were subsequently stratified into low-risk and high-risk groups. The practicality of the NRG signature was validated using K–M analysis, receiver operating characteristic (ROC), principal component analysis (PCA), and a heatmap. Univariate and multivariate Cox analyses were performed to determine whether the signature was an independent risk factor for patient prognosis. A nomogram was constructed using the ‘rms’ package, incorporating clinicopathological parameters and NRG signatures. Calibration plots were then used to assess the consistency between the predicted survival events and actual outcomes.

2.5. Immune landscape analysis

The ESTIMATE algorithm was utilized to evaluate the immune score, stromal score, and ESTIMATE score in the tumor microenvironment (TME). The single sample GSEA (ssGSEA) was conducted to investigate the activity of immune cells, immune functions, and immune pathways for each sample. In addition, the expression levels of major histocompatibility complex (MHC) and common immune checkpoint molecules were compared between the identified subgroups. Furthermore, the Tumor Immune Dysfunction and Exclusion (TIDE) score was calculated for patients with GC using the TIDE database. The TIDE score has been correlated with an inferior response to immune checkpoint blockade therapy and a shorter OS. This score helps identify patients who may benefit more from immune checkpoint therapy.

2.6. Chemotherapy response and drug sensitivity

To investigate the correlation between drug sensitivity and the NRG signature, the R package ‘pRRophetic’ was utilized. The half-maximal inhibitory concentration (IC50) values of low- and high-risk GC patients were compared using Wilcoxon signed-rank tests.

2.7. Deep learning models based on H&E stained WSIs

This study employed deep learning techniques to analyze the morphological characteristics of tumor tissues and infer potential molecular alterations in patients. According to our previous research [18], convolutional neural networks with deep residual learning (ResNet18 and ResNet50) were employed to classify GC patients based on the NRG signature using tumor pathological images stained with H&E. ResNet18 was used for tumor extraction. The ResNet50 model was used to build a high-risk population identification model. Pathological images were obtained from 130 patients in the TCGA-STAD cohort, which were randomly divided into training (91 cases) and validation (39 cases) subsets at a 7:3 ratio.

WSIs, which were too large to process directly, were divided into 224 × 224 pixels using the OpenSlide package to facilitate subsequent analyses. Non-tissue white background was eliminated using Otsu’s approach. To mitigate inconsistencies across different laboratories, a normalization technique was employed to convert all images to a standardized reference color space. The Adam optimizer was utilized to train the models, and the process involved 100 iterations.

2.8. Identification of hub genes of the NRG signature

The protein-protein interaction (PPI) network of the NRG signature genes was constructed using the STRING database, and the CytoHubba plugin of Cytoscape software (version 3.9.1) was utilized to identify the hub genes.

2.9. Real-time quantitative real-time PCR (RT-qPCR)

We used RT-qPCR to detect the relative mRNA expression levels of hub genes in both GC tissues and controls. Primer sequences were designed using the Primer-5 software and are provided in Supplementary Table 2. Total RNA was extracted from GC tissue samples using an RNA Easy Fast Animal Tissue Total RNA Extraction Kit (DP451, Tiangen Biotech Co., Ltd., China) according to the manufacturer’s instructions. cDNA was synthesized using MonScript™ RTIII All-in-One Mix with a dsDNase Kit (MR05101, Monad Biotech Co., Ltd., China). RT-qPCR was performed using a SYBR PCR Kit (Q711-02, Nanjing Vazyme Biotech Co., Ltd., China). The reaction program included: (1) 95 °C for 1 min, (2) 95 °C for 15 s, and 65 °C for 30 s, repeated for 40 cycles. The 2−ΔΔCt method was used to calculate the relative expression level of each target mRNA.

2.10. Single-cell RNA sequencing data analysis

Single-cell dataset analysis was conducted using Scanpy (version 1.9.1) [19]. Preprocessing was performed according to standard procedures detailed on the Scanpy website (https://scanpy-tutorials.readthedocs.io/en/latest/pbmc3k.html). The expression data were then normalized and logarithmized, and the effects of total counts per cell were regressed out, followed by scaling to the same range. After preprocessing, we reduced the dimensionality of the data by performing PCA. Subsequently, the UMAP function was used to further reduce the dimensionality of the merged datasets. Finally, the cells were clustered based on the neighborhood graph using the Leiden clustering method.

2.11. Statistical analyses

Differences between groups were compared using the Wilcoxon test, and correlation coefficients were calculated using Spearman’s analysis. K–M survival analyses with log-rank tests were used to evaluate significant differences in OS between the groups. The performance of each model was evaluated by calculating the area under the ROC curve (AUC). The percentile bootstrap method was used to present 95% confidence intervals (CIs). Statistical significance was defined as a p-value <0.05. All statistical analyses were conducted using R version 4.1.0 or Python 3.8.13.

3. Results

3.1. Identification of NRG clusters and prognostic differences

The ‘ConsensusClusterPlus’ R package was used to perform consistent cluster analysis in the TCGA-STAD cohort, based on the expression of 371 NRGs. The clustering variable (k) was set from 2 to 10, and the analysis showed that when k = 2, there was the most significant correlation within groups and a low correlation between the groups (Figures 1A–D). Consequently, all the samples were divided into two clusters: C1 (n = 226) and C2 (n = 145). Samples with available survival information were used to explore the survival differences between clusters. K–M analysis revealed that patients in cluster C2 had better OS than those in cluster C1 (p = 0.032), and the median OS time in C2 was also longer than that in C1 (Figure 1E). Additionally, a heatmap was generated to display the patients’ clinical information (histological grade, pathological stage, T stage, N stage, M stage, age, and gender) between the two clusters. Notably, the heatmap revealed significant differences in the histological grade, pathological stage, and T stage between C1 and C2 (Figure 1F).

Figure 1.

Figure 1.

Open in a new tab

Cluster analysis of TCGA-STAD patients based on NRGs. (A–D) A total of 371 GC patients were divided into two distinct clusters according to the consensus clustering matrix. (E) K–M survival analysis between different cluster groups. (F) Heatmap displaying the distribution of clinicopathological variables between different cluster groups (*p < 0.05; **p < 0.01).

3.2. TME, ssGSEA, and HLA analysis based on NRG clusters

The ESTIMATE algorithm analysis revealed that C1 had increased immune score, stromal score, and ESTIMATE score, while C2 showed the opposite trend (Figures 2A–C). Additionally, ssGSEA was conducted to assess immune cell infiltration in both clusters. The analysis revealed 23 immune cell types that exhibited different levels of infiltration between clusters. The boxplot (Figure 2D) showed that C1 had significant infiltration of most immune cells, including activated B cells, activated CD8+ T cells, activated dendritic cells, mast cells, macrophages, and myeloid-derived suppressor cells (MDSCs). In contrast, C2 showed higher expression levels only in activated CD4+ T cells, natural killer cells, neutrophils, and type.17. T. helper cells. Intriguingly, C1 also exhibited higher expression levels of MHC molecules (Figure 2E).

Figure 2.

Figure 2.

Open in a new tab

Comparison of immune profiles between the two distinct clusters. (A–C) Distribution of immune score, stromal score, and ESTIMATE score between C1 and C2. (D) Discrepancies of immune cell infiltration between C1 and C2. (E) Gene expression of HLA gene sets between clusters (*p < 0.05; **p < 0.01; ***p < 0.001).

3.3. Construction and validation of the NRG signature

Univariate Cox regression analysis was performed using the DEGs from the clusters to identify genes that were significantly associated with survival, resulting in 218 prognostic genes (p < 0.01) for further analysis. To refine the prognostic gene set, LASSO regression was employed, and 13 prognostic genes were selected. Subsequently, these 13 NRG-related prognostic genes were used to construct an NRG signature (Figures 3A–C). The NRG signature formula was as follows: Risk score = (0.058 × PRTG expression) + (0.114 × BPI expression) + (0.012 × RGS2 expression) + (0.034 × SLC7A2 expression) + (0.093 × SERPINE1 expression) + (0.019 × CYTL1 expression) + (0.0002 × CXCR4 expression) + (0.024 × RNASE3 expression) + (0.036 × MATN3 expression) + (0.004 × GPX3 expression) + (0.081 × GDF6 expression) + (0.017 × APOD expression) + (0.119 × CGB5 expression). Patients were stratified into two groups, namely high-risk and low-risk groups, based on the median risk score derived from the NRG signature.

Figure 3.

Figure 3.

Open in a new tab

Construction and validation of the NRG signature based on NRG-related clusters. (A) Forest plot showing the 13 genes selected in the signature through LASSO analysis. (B) Selection of the best tuning parameters (logλ) for 10-fold cross-validation. (C) 13 genes to construct the signature model. K–M curves of high- and low-risk patients in the training set (D) and test set (E). ROC curves of high- and low-risk patients in the training set (F) and test set (G). PCA (H) and t-SNE (I) analysis between the high- and low-risk groups in the training set and test set. Ranking points (J) and scatter plots (K) showing NRG score distributions and patient survival status in the training set and test set.

The K–M curves provided clear evidence that the high-risk group exhibited a significantly worse prognosis than the low-risk group in both training and test sets (p < 0.001; Figures 3D,E). The 1-year, 3-year, and 5-year OS represented by the AUC values of the NRG signature in the training set were 0.633, 0.709, and 0.752, respectively, reflecting robust sensitivity and specificity (Figure 3F). Similarly, in the GSE84437 validation cohort, the AUC values for 1-year, 3-year, and 5-year OS were 0.624, 0.648, and 0.635, respectively, further confirming the predictive performance of the NRG signature (Figure 3G). In addition, PCA and t-distributed stochastic neighbor embedding (t-SNE) analysis showed that this NRG signature achieved distinct clustering of patients, indicating a good resolution (Figures 3H,I). Furthermore, the risk score distribution plot and survival state heatmap indicated that the high-risk group had a greater proportion of death, accompanied by reduced survival times as the risk score increased (Figures 3J,K).

3.4. Independent prognostic value of the NRG signature and construction of the nomogram

A heatmap was constructed to explore the correlation between clinicopathological features and the NRG signature. Interestingly, the heatmap revealed that the NRG signature was associated with histological grade, pathological stage, and T stage, but showed no significant correlation with other clinical features, such as age and gender (Figure 4A). Subsequently, univariate and multivariate Cox regression analyses were performed in both TCGA and GEO cohorts. These findings indicated that the NRG signature was an independent prognostic factor for patients with GC (Figures 4B,C).

Figure 4.

Figure 4.

Open in a new tab

Establishment and assessment of the nomogram for GC survival prediction. (A) Heatmap for the NRG signature based on the risk groups and clinicopathological manifestation. Univariate and multivariate Cox regression analyses showed that the NRG signature is an independent prognostic factor affecting the prognosis of GC patients in the training set (B) and test set (C). (D) Development of the nomogram combining NRG risk score and other clinicopathological parameters to predict 1-, 3-, and 5-year survival. (E) DCA evaluating the 1-year OS in the TCGA-STAD set. (F) Calibration curves displaying the predictions of the established nomogram for 1-, 3-, and 5-year OS. (G) ROC curves comparing the predictive capability of age, gender, grade, stage, NRG signature, and the nomogram in predicting OS (*p < 0.05; **p < 0.01; ***p < 0.001).

To create a more comprehensive tool for predicting 1-, 3-, and 5-year survival rates in GC, a nomogram incorporating the NRG signature along with other clinicopathological features was constructed (Figure 4D). Time-dependent ROC and decision curve analysis (DCA) were used to evaluate the sensitivity and specificity of the nomogram for prognosis. DCA results indicated that the nomogram was significantly superior to the use of a single independent predictor (Figure 4E). The nomogram incorporating multiple clinical factors exhibited a higher net benefit in predicting the prognosis of GC patients. Additionally, the calibration plot of the nomogram showed excellent consistency between the predicted and actual observation probabilities (Figure 4F). Furthermore, the ROC analysis showed that the nomogram had the highest AUC (0.691) compared to other individual factors (signature 0.630, age 0.583, gender 0.542, histological grade 0.565, and pathological stage 0.603), indicating that the nomogram provided the most stable and accurate prediction of survival outcomes in GC patients (Figure 4G).

3.5. Functional analysis between the high- and low-risk groups based on the NRG signature

To explore the functional disparities between the high-risk and low-risk groups stratified by the NRG signature, GO and KEGG analyses were performed. GO enrichment analysis (Supplementary Figure S1A) revealed that DEGs between the two groups were primarily involved in extracellular matrix organization, extracellular structure organization, regulation of cytosolic calcium ion concentration, cell-substrate adhesion, and other biological processes. Additionally, KEGG results (Supplementary Figure S1B) showed that DEGs were significantly enriched in the PI3K-Akt signaling pathway, focal adhesion, cytokine-cytokine receptor interaction, and ECM receptor interaction TGF-beta signaling pathway.

3.6. Association of the NRG signature with clinical immune characteristics and immune checkpoints

The ESTIMATE analysis showed that the high-risk group exhibited higher immune, stromal, and ESTIMATE scores, while the low-risk group demonstrated the opposite results (Figures 5A–C). Additionally, ssGSEA revealed higher immune cell infiltration and immune pathway activation in the high-risk group (Figures 5D,E). Furthermore, this study investigated the relationship between 13 signature genes and immune cells. The correlation heatmap (Figure 5F) demonstrated a positive correlation between these genes and the immune cells. In addition, the expression levels of common checkpoint genes between the high- and low-risk groups were compared. A boxplot revealed statistically significant differences in checkpoint gene expression between the two groups (Figure 5G), except for CD274, PDCD1, and CTLA4 (p > 0.05). These findings indicated that the NRG signature may serve as a useful tool for predicting immune response and providing vital information for optimizing personalized immunotherapy.

Figure 5.

Figure 5.

Open in a new tab

Assessment of the tumor microenvironment and immune checkpoint genes in different groups. (A–C) Comparison of immune score, stromal score, and ESTIMATE score between the low- and high-risk groups. (D,E) Discrepancies of immune cell infiltration and expression of immune-related pathways between the low- and high-risk groups. (F) Correlation analysis between the signature genes and immune cells. (G) Differential expression analysis of immune checkpoint genes between the low- and high-risk groups (*p < 0.05; **p < 0.01; ***p < 0.001).

3.7. Prediction of drug sensitivity and immunotherapy response based on the NRG signature

We evaluated the drug sensitivity in patients with GC. These results indicated significant differences in chemotherapeutic sensitivity between the low- and high-risk groups. Specifically, the low-risk group (Figures 6A–F) showed higher sensitivity to ABT.888 (p = 2e-06), BIBW2992 (p = 1.8e-09), Gefitinib (p = 4e-05), Metformin (p = 1.5e-05), SB590885 (p = 0.00017), and GW.441756 (p = 6.5e-05). In contrast, the high-risk group (Figures 6G–K) demonstrated higher sensitivity to Rapamycin (p = 6.5e-05), Shikonin (p = 6.2e-06), Sunitinib (p = 0.00046), Imatinib (p = 6.9e-13), and Dasatinib (p = 5.3e-10). Furthermore, the TIDE score (Figure 6L) was used to assess the differences in immunotherapy response between the high- and low-risk groups. The results indicated a significantly higher TIDE score in the high-risk group compared to the low-risk group, suggesting more immune dysfunctions and exclusion mechanisms in the tumor microenvironment, which could contribute to a poor response to immunotherapy.

Figure 6.

Figure 6.

Open in a new tab

NRG signature predicts chemotherapy and immunotherapy response. (A–F) The signature identified that low risk scores were associated with lower IC50 values for chemotherapeutics, such as (A) ABT.888, (B) BIBW2992, (C) Gefitinib, (D) Metformin, (E) SB590885, (F) GW.441756, whereas high risk scores were related to lower IC50 values for (G) Rapamycin, (H) Shikonin, (I) Sunitinib, (J) Imatinib, and (K) Dasatinib treatment. (L) Differences in TIDE score between the high- and low-risk groups.

3.8. Prediction of the NRG signature from H&E WSIs using deep learning model

First, we preprocessed the WSIs by dividing them into 224 × 224 pixel tiles using the OpenSlide package, ensuring an appropriate size for subsequent analysis. Subsequently, we utilized the ResNet18 architecture for tumor extraction, which is a critical step in isolating tumor regions from WSIs. This enabled us to focus our analysis specifically on tumor-related features. Following tumor extraction, we developed a more advanced deep learning model using ResNet50 to predict the NRG signature. The ResNet50 model, with its deeper architecture, was better suited for capturing complex and subtle patterns within the images, thereby enhancing the predictive capabilities of our model. The results demonstrated that the deep learning model effectively distinguished GC patients into high- and low-risk groups by leveraging morphological variances (AUC = 0.6322, 95%CI = 0.5747–0.6764). The corresponding processes and ROC curve are shown in Figure 7. Overall, our comprehensive approach, combining tile-based WSI preprocessing, tumor extraction with ResNet18, and the NRG signature prediction with ResNet50, has shown promising results in the classification of patients with GC.

Figure 7.

Figure 7.

Open in a new tab

Overview of the deep learning model. (A) The WSIs of each patient were obtained from the TCGA database. These WSIs were divided into 224 × 224 pixels using the OpenSlide package. (B) The extracted sub-images were utilized as input to train the ResNet18 model for automatic extraction of tumor tissues. A ResNet50 model was constructed to predict the NRG signature (C), and its performance was evaluated using a ROC curve (D).

3.9. Differential expression analysis and prognostic analysis of the signature genes

To further verify the credibility of the NRG signature, we first conducted differential gene expression analysis using TCGA data. Based on the analytical results (Supplementary Figure S2A), the following signature genes were upregulated in GC tissues: PRTG, BPI, SLC7A2, SERPINE1, CXCR4, RNASE3, MATN3, and CGB5. Conversely, RGS2, CYTL1, GPX3, GDF6, and APOD were upregulated in the adjacent non-tumor tissues. To assess the association between the signature genes and the prognosis of patients with GC, we conducted a survival analysis using GEPIA2 (http://gepia2.cancer-pku.cn/#index). K–M survival analysis (Supplementary Figures S2B–N) demonstrated a significant association between all 13 genes and OS in patients with GC.

Furthermore, we identified hub signature genes and validated their expression at mRNA level. A PPI network of the signature genes was constructed using the STRING database (Figure 8A). Subsequently, we used the MNC, EPC, and Degree algorithms based on CytoHubba to identify the top five hub genes, including SERPINE1, CXCR4, GPX3, APOD, and MATN3. To better characterize the expression levels of these genes, we collected 26 pairs of adjacent normal gastric tissue and GC specimens. As shown in Figures 8B–F, compared to adjacent normal gastric tissue specimens, the expression level of SERPINE1 was significantly upregulated in GC specimens, whereas the expression levels of GPX3 and APOD were significantly downregulated.

Figure 8.

Figure 8.

Open in a new tab

The PPI network and hub gene expression analysis. (A) The PPI network of signature genes. (B–F) The mRNA expression levels of hub genes (*p < 0.05; **p < 0.01; ***p < 0.001).

3.10. Utilization of single-cell RNA sequencing data to identify the expression of hub signature genes

After dimension reduction and clustering, we identified 10 distinct major clusters representing the epithelial, immune, and endothelial populations (Figure 9A). Upon comparing the gene expression profiles of tumor and normal cells, we found that CXCR4, RGS2, and SERPINE1 were highly expressed in tumor cells, whereas GPX3 and APOD were highly expressed in normal cells (Figure 9B). To investigate the distribution of hub signature genes across these clusters, we first drew feature plots and then constructed violin plots to display gene expression (Figures 9C–G). Our analysis revealed that CXCR4 was predominantly expressed in CD8T+ cells, B cells, and CD4 T cells; GPX3 was mainly expressed in endothelial cells and stromal cells, SERPINE1 was mainly expressed in endothelial cells, and APOD was primarily expressed in stromal cells. The remaining signature genes are presented in Supplementary Figure S3.

Figure 9.

Figure 9.

Open in a new tab

Identification and annotation of cell clusters. (A) Leiden clustering analysis identifies 10 distinct cell clusters. (B) The expression of signature genes in both tumor and normal cells. (C–G) The cell localization of and expression patterns of hub genes.

4. Discussion

GC is highly heterogeneous, necessitating the identification of multigene biomarkers for prognosis and personalized treatment. This study specifically focuses on the association between NRGs and GC, offering a complementary investigation to existing research. Diverse forms of non-apoptotic RCD could effectively bypass or overcome the resistance of tumor cells to apoptosis, offering alternative death pathways when the apoptotic pathway is defective, thereby significantly enhancing anti-cancer efficacy. We developed a prognostic risk signature comprising 13 NRGs and demonstrated that the NRG signature could independently predict the OS of GC patients, as validated by the GEO database. Furthermore, we presented a deep learning model with good performance in predicting the NRG signature in GC patients using H&E-stained WSIs.

Targeting non-apoptotic RCD has received significant attention in the field of anti-tumor therapy owing to its impact on cancer development and response to treatment, in contrast to apoptosis [4]. By focusing on NRG expression profiles, we identified two distinct clusters (C1 and C2) in GC patients with significant differences in clinical characteristics, immune microenvironment, and prognosis. We found that the clinical features, prognosis, and immune cell infiltration differed between the two clusters, suggesting that NRGs may be broadly involved in the mechanisms underlying GC progression. Based on the DEGs between the two clusters, a 13-gene NRG signature was constructed and validated as an independent prognostic factor for patients with GC. ROC analysis indicated that the accuracy of the NRG signature for prognostic prediction in GC patients was favorable, with AUC values of 0.633, 0.709, and 0.752 at 1, 3, and 5 years, respectively. We also designed a nomogram combining the NRG signature and relevant clinical features, which exhibited exceptional predictive ability with reliable calibration. Moreover, patients with high-risk scores exhibited lower sensitivity to chemotherapy and targeted therapy drugs, including Sunitinib [20], Shikonin [21], Dasatinib [22], Rapamycin [23], and Imatinib [24], potentially explaining their inferior survival outcomes. Generally, our results suggest that the NRG signature, based on non-apoptotic RCD, is a significant prognostic marker that could effectively predict prognosis, certain clinical features, and drug response of GC patients, suggesting its promising clinical value.

Previous studies have suggested that a higher stromal population is linked to tumor progression, as it can alter anti-tumor immunity and influence responsiveness to immunotherapy [25,26]. Some studies have also proposed that in immunologically active hosts during tumor development, less immunogenic cancer cells might be favored, and immunosuppressive networks could be established to evade anti-tumor immune responses [27]. Our study revealed a close relationship between NRGs and infiltration levels of different immune cells. In contrast to the low-risk group, the high-risk group displayed elevated stromal, immune, and ESTIMATE scores. Similarly, we found elevated expression levels of immunosuppressive cells, including MDSCs, mast cells, regulatory T cells (Tregs), and T follicular helper cells (Tfh), in patients with high NRG scores. These cells have been reported to promote immunosuppression and reduce cancer antigen levels and immune reactivity [28–30]. Moreover, studies have linked checkpoint genes to immunosuppressive features [31,32]. Our findings revealed higher checkpoint gene expression levels and TIDE scores in the high-risk group, suggesting that the risk score effectively distinguished the extent of immune escape in patients and provided insights into the underlying reasons for the observed poor survival rates in this group. It is evident from our results that patients with a lower risk score exhibited a more favorable response to immunotherapy and were less likely to experience immune escape. Given these significant observations, it is reasonable to consider using NRG signature-based classification as a basis for selecting appropriate therapeutic strategies, particularly for immunotherapy, in GC patients.

In recent years, a growing number of researchers have developed cancer prognostic models that utilize novel mechanisms of cell death. However, their practical implementation in a clinical setting is lacking. Bridging the gap between pathomics and genomics represents a promising new direction as it establishes a vital connection between model performance, theoretical concepts, and real-world applications. This approach is essential for the comprehensive validation of model performance prior to clinical implementation. H&E-stained histopathological images have been proven to provide crucial information for aiding clinical decision-making among numerous cancers [33,34]. With the advent of deep learning and the availability of a large number of histological slides, there is a unique opportunity to reassess traditional techniques for predicting patient diagnosis and prognosis [35]. In this study, we developed a deep learning model based on H&E-stained WSIs to facilitate the NRG signature-based classification of GC, which achieved an AUC of 0.679, demonstrating its potential clinical utility. This approach eliminates the need for costly and limited availability of the whole-transcriptome sequencing data. Our model offers a reliable means to classify GC patients through pathological images and the NRG signature, and appropriate treatment options can be selected to administer the most suitable drug for the patient’s condition. Compared to the previous conventional model, the integration of pathologic patterns into the current model broadens its applicability and renders it more than mere speculation in predicting patient prognosis. This approach allows the identification of distinct patient differences from a pathological perspective.

From this signature, we further identified and verified five hub genes, including SERPINE1, CXCR4, APOD, GPX3, and MATN3. These genes have been identified in various types of cancer and are associated with tumor progression and prognosis. SERPINE1 is highly expressed in tumors and is associated with tumor prognosis [36]. In GC, high levels of SERPINE1 can promote proliferation, invasion, migration, and angiogenesis of tumor cells, resulting in poorer patient outcomes [37]. CXCR4 belongs to the GPCR family and has been implicated in various cancers. Elevated CXCR4 expression in cancer tissue has been inversely correlated with GC prognosis [38,39], and inhibition of CXCR4 expression significantly impairs the proliferation, migration, and invasion of GC cells [40]. APOD expression is downregulated in several types of cancers, such as hepatocellular carcinoma and colorectal cancer [41,42]. Wang et al. suggested that APOD mRNA expression was significantly decreased in GC samples, with higher levels of APOD expression indicative of poor prognosis [43]. GPX3 may play a dual role in various cancers [44]. Hu et al. found that the expression of GPX3 was downregulated in GC but was positively correlated with poor outcomes [45], which is consistent with our results. However, Khan et al. found that patients with GC with higher GPX3 expression tended to have a poor prognosis [46]. MATN3 is a member of the extracellular matrix protein family. Aberrant expression patterns of MATN3 have been reported in pancreatic ductal adenocarcinoma and osteosarcoma [47,48]. Elevated expression of the MATN3 protein was observed in GC patients, and lower levels of MATN3 were found to be negatively correlated with survival time [49]. Overall, the substantial impact of these genes on cancer enhanced the reliability of the signature. Further exploration is necessary to clarify the specific mechanisms by which these genes exert their effects in GC.

However, this study has several limitations that need to be acknowledged. First, the data is primarily from TCGA and GEO databases, necessitating validation in large, multi-center clinical trials and more independent datasets, along with a prospective study, to confirm the predictive significance. Second, further experimental validation is required to elucidate the role of the NRG signature in the development of GC. Third, given the heterogeneity of GC, data from distinct datasets are likely to affect the overall findings. Lastly, while our study provides insight into the possible relationship between the NRG signature and immune status, the exact regulatory mechanism underlying this relationship remains unclear.

5. Conclusion

In this study, we established an NRG risk signature based on non-apoptotic RCD-related genes and used it to classify GC into high- and low-risk groups. The different NRG risk groups exhibited distinct clinical characteristics, immune cell profiles, and therapy responses. Furthermore, a nomogram based on the NRG signature was constructed to predict the survival of patients with GC. Finally, a deep learning model was developed to facilitate NRG-signature-based classification of GC from pathological images. The results of our study suggest that the NRG signature could serve as a valuable tool for guiding clinical decision-making and personalized therapy strategies for patients with GC.

Supplementary Material

Supplemental Material
IANN_A_2426758_SM3325.zip (10.3MB, zip)

Acknowledgments

There are no additional acknowledgments of the financial, material, or author support.

Funding Statement

This work was funded by a project from the Science and Technology Department of Liaoning Province to support the high-quality development of China Medical University (2023JH2/20200094) and the Basic Research Project of Liaoning Provincial Department of Education (LJKMZ20222209).

Ethical approval

In this study, all data were downloaded from TCGA and GEO databases. There were no restrictions on the use of TCGA and GEO data for research and analysis purposes. All data can be downloaded and used freely and do not require an ethics statement.

All patients included in the validation experiments for this study provided consent forms and ethical approval was obtained from the Ethics Committee of the First Hospital of China Medical University (Ethical Approval Number: [2023] 2023-446-2).

Author contributions

JJ, XZ, and QS were involved in the study conception and design. ZY, TL, and QS were responsible for collecting clinicopathological data from patients. XZ, ZW, and TL analyzed and interpreted the data. QS drafted the manuscript. JJ and XZ revised the manuscript accordingly. JJ and XZ supervised the study and provided administrative, technical, and material support throughout the study. All authors have reviewed and approved the final manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, Jingjing Jing, upon reasonable request.

References

  • 1.Li K, Zhang A, Li X, et al. Advances in clinical immunotherapy for gastric cancer. Biochim Biophys Acta Rev Cancer. 2021;1876(2):188615. doi: 10.1016/j.bbcan.2021.188615. [DOI] [PubMed] [Google Scholar]
  • 2.Kim SW, Roh J, Lee HS, et al. Expression of the immune checkpoint molecule V-set immunoglobulin domain-containing 4 is associated with poor prognosis in patients with advanced gastric cancer. Gastric Cancer. 2021;24(2):327–340. doi: 10.1007/s10120-020-01120-1. [DOI] [PubMed] [Google Scholar]
  • 3.Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209–249. doi: 10.3322/caac.21660. [DOI] [PubMed] [Google Scholar]
  • 4.Koren E, Fuchs Y.. Modes of regulated cell death in cancer. Cancer Discov. 2021;11(2):245–265. doi: 10.1158/2159-8290.CD-20-0789. [DOI] [PubMed] [Google Scholar]
  • 5.Gao W, Wang X, Zhou Y, et al. Autophagy, ferroptosis, pyroptosis, and necroptosis in tumor immunotherapy. Signal Transduct Target Ther. 2022;7(1):196. doi: 10.1038/s41392-022-01046-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Badgley MA, Kremer DM, Maurer HC, et al. Cysteine depletion induces pancreatic tumor ferroptosis in mice. Science. 2020;368(6486):85–89. doi: 10.1126/science.aaw9872. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Gu Z, Liu T, Liu C, et al. Ferroptosis-strengthened metabolic and inflammatory regulation of tumor-associated macrophages provokes potent tumoricidal activities. Nano Lett. 2021;21(15):6471–6479. doi: 10.1021/acs.nanolett.1c01401. [DOI] [PubMed] [Google Scholar]
  • 8.Qiao Y, Choi JE, Tien JC, et al. Autophagy inhibition by targeting PIKfyve potentiates response to immune checkpoint blockade in prostate cancer. Nat Cancer. 2021;2(9):978–993. doi: 10.1038/s43018-021-00237-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Xie Y, Zhu S, Zhong M, et al. Inhibition of aurora kinase A induces necroptosis in pancreatic carcinoma. Gastroenterology. 2017;153(5):1429–1443.e1425. doi: 10.1053/j.gastro.2017.07.036. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Niazi MKK, Parwani AV, Gurcan MN.. Digital pathology and artificial intelligence. Lancet Oncol. 2019;20(5):e253–e261. doi: 10.1016/S1470-2045(19)30154-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Jang HJ, Lee A, Kang J, et al. Prediction of genetic alterations from gastric cancer histopathology images using a fully automated deep learning approach. World J Gastroenterol. 2021;27(44):7687–7704. doi: 10.3748/wjg.v27.i44.7687. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Lee SH, Jang HJ.. Deep learning-based prediction of molecular cancer biomarkers from tissue slides: a new tool for precision oncology. Clin Mol Hepatol. 2022;28(4):754–772. doi: 10.3350/cmh.2021.0394. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Jang HJ, Lee A, Kang J, et al. Prediction of clinically actionable genetic alterations from colorectal cancer histopathology images using deep learning. World J Gastroenterol. 2020;26(40):6207–6223. doi: 10.3748/wjg.v26.i40.6207. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Tran KA, Kondrashova O, Bradley A, et al. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med. 2021;13(1):152. doi: 10.1186/s13073-021-00968-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Wen Z, Wang S, Yang DM, et al. Deep learning in digital pathology for personalized treatment plans of cancer patients. Semin Diagn Pathol. 2023;40(2):109–119. doi: 10.1053/j.semdp.2023.02.003. [DOI] [PubMed] [Google Scholar]
  • 16.Yamashita R, Long J, Longacre T, et al. Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. Lancet Oncol. 2021;22(1):132–141. doi: 10.1016/S1470-2045(20)30535-0. [DOI] [PubMed] [Google Scholar]
  • 17.Raczkowski L, Pasnik I, Kukielka M, et al. Deep learning-based tumor microenvironment segmentation is predictive of tumor mutations and patient survival in non-small-cell lung cancer. BMC Cancer. 2022;22(1):1001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Wei Z, Zhao X, Chen J, et al. Deep learning-based stratification of gastric cancer patients from hematoxylin and eosin-stained whole slide images by predicting molecular features for immunotherapy response. Am J Pathol. 2023;193(10):1517–1527. doi: 10.1016/j.ajpath.2023.06.004. [DOI] [PubMed] [Google Scholar]
  • 19.Wolf FA, Angerer P, Theis FJ.. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 2018;19(1):15. doi: 10.1186/s13059-017-1382-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Yi JH, Lee J, Lee J, et al. Randomised phase II trial of docetaxel and sunitinib in patients with metastatic gastric cancer who were previously treated with fluoropyrimidine and platinum. Br J Cancer. 2012;106(9):1469–1474. doi: 10.1038/bjc.2012.100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Lee MJ, Kao SH, Hunag JE, et al. Shikonin time-dependently induced necrosis or apoptosis in gastric cancer cells via generation of reactive oxygen species. Chem Biol Interact. 2014;211:44–53. doi: 10.1016/j.cbi.2014.01.008. [DOI] [PubMed] [Google Scholar]
  • 22.Choi KM, Cho E, Bang G, et al. Activity-based protein profiling reveals potential dasatinib targets in gastric cancer. Int J Mol Sci. 2020;21(23):9276. doi: 10.3390/ijms21239276. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Lang SA, Gaumann A, Koehl GE, et al. Mammalian target of rapamycin is activated in human gastric cancer and serves as a target for therapy in an experimental model. Int J Cancer. 2007;120(8):1803–1810. doi: 10.1002/ijc.22442. [DOI] [PubMed] [Google Scholar]
  • 24.Mayr M, Becker K, Schulte N, et al. Phase I study of imatinib, cisplatin and 5-fluoruracil or capecitabine in advanced esophageal and gastric adenocarcinoma. BMC Cancer. 2012;12(1):587. doi: 10.1186/1471-2407-12-587. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Turley SJ, Cremasco V, Astarita JL.. Immunological hallmarks of stromal cells in the tumour microenvironment. Nat Rev Immunol. 2015;15(11):669–682. doi: 10.1038/nri3902. [DOI] [PubMed] [Google Scholar]
  • 26.Demirkol Canli S, Uner M, Kucukkaraduman B, et al. A novel gene list identifies tumors with a stromal-mesenchymal phenotype and worse prognosis in ­gastric cancer. Cancers. 2023;15(11):3035. doi: 10.3390/cancers15113035. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Dunn GP, Bruce AT, Ikeda H, et al. Cancer immunoediting: from immunosurveillance to tumor escape. Nat Immunol. 2002;3(11):991–998. doi: 10.1038/ni1102-991. [DOI] [PubMed] [Google Scholar]
  • 28.Zou W, Chen L.. Inhibitory B7-family molecules in the tumour microenvironment. Nat Rev Immunol. 2008;8(6):467–477. doi: 10.1038/nri2326. [DOI] [PubMed] [Google Scholar]
  • 29.Joshi NS, Akama-Garren EH, Lu Y, et al. Regulatory T cells in tumor-associated tertiary lymphoid structures suppress anti-tumor T cell responses. Immunity. 2015;43(3):579–590. doi: 10.1016/j.immuni.2015.08.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Veglia F, Sanseviero E, Gabrilovich DI.. Myeloid-derived suppressor cells in the era of increasing myeloid cell diversity. Nat Rev Immunol. 2021;21(8):485–498. doi: 10.1038/s41577-020-00490-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Thommen DS, Schumacher TN.. T cell dysfunction in cancer. Cancer Cell. 2018;33(4):547–562. doi: 10.1016/j.ccell.2018.03.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Batlle E, Massagué J.. Transforming growth factor-β signaling in immunity and cancer. Immunity. 2019;50(4):924–940. doi: 10.1016/j.immuni.2019.03.024. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Jiang Y, Yang M, Wang S, et al. Emerging role of deep learning-based artificial intelligence in tumor pathology. Cancer Commun. 2020;40(4):154–166. doi: 10.1002/cac2.12012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Bera K, Schalper KA, Rimm DL, et al. Artificial intelligence in digital pathology – new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16(11):703–715. doi: 10.1038/s41571-019-0252-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Zheng X, Wang R, Zhang X, et al. A deep learning model and human-machine fusion for prediction of EBV-associated gastric cancer from histopathology. Nat Commun. 2022;13(1):2790. doi: 10.1038/s41467-022-30459-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Liu X, Wu J, Zhang D, et al. Identification of potential key genes associated with the pathogenesis and prognosis of gastric cancer based on integrated bioinformatics analysis. Front Genet. 2018;9:265. doi: 10.3389/fgene.2018.00265. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Yang JD, Ma L, Zhu Z.. SERPINE1 as a cancer-promoting gene in gastric adenocarcinoma: facilitates tumour cell proliferation, migration, and invasion by regulating EMT. J Chemother. 2019;31(7–8):408–418. doi: 10.1080/1120009X.2019.1687996. [DOI] [PubMed] [Google Scholar]
  • 38.Wen F, Lu X, Huang W, et al. Characteristics of immunophenotypes and immunological in tumor microenvironment and analysis of immune implication of CXCR4 in gastric cancer. Sci Rep. 2022;12(1):5720. doi: 10.1038/s41598-022-08622-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Xue LJ, Mao XB, Ren LL, et al. Inhibition of CXCL12/CXCR4 axis as a potential targeted therapy of advanced gastric carcinoma. Cancer Med. 2017;6(6):1424–1436. doi: 10.1002/cam4.1085. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Xue S, Ma M, Bei S, et al. Identification and validation of the immune regulator CXCR4 as a novel promising target for gastric cancer. Front Immunol. 2021;12:702615. doi: 10.3389/fimmu.2021.702615. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Utsunomiya T, Ogawa K, Yoshinaga K, et al. Clinicopathologic and prognostic values of apolipoprotein D alterations in hepatocellular carcinoma. Int J Cancer. 2005;116(1):105–109. doi: 10.1002/ijc.20986. [DOI] [PubMed] [Google Scholar]
  • 42.Ogawa K, Utsunomiya T, Mimori K, et al. Genomic screens for genes upregulated by demethylation in colorectal cancer: possible usefulness for clinical application. Int J Oncol. 2005;27(2):417–426. doi: 10.3892/ijo.27.2.417. [DOI] [PubMed] [Google Scholar]
  • 43.Wang J, Liu D, Wang Q, et al. Identification of basement membrane-related signatures in gastric cancer. Diagnostics. 2023;13(11):1844. doi: 10.3390/diagnostics13111844. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Chang C, Worley BL, Phaëton R, et al. Extracellular glutathione peroxidase GPx3 and its role in cancer. Cancers. 2020;12(8):2197. doi: 10.3390/cancers12082197. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Hu Q, Chen J, Yang W, et al. GPX3 expression was down-regulated but positively correlated with poor outcome in human cancers. Front Oncol. 2023;13:990551. doi: 10.3389/fonc.2023.990551. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Khan M, Lin J, Wang B, et al. A novel necroptosis-related gene index for predicting prognosis and a cold tumor immune microenvironment in stomach adenocarcinoma. Front Immunol. 2022;13:968165. doi: 10.3389/fimmu.2022.968165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Ding J, Liu Y, Lai Y.. Identifying MMP14 and COL12A1 as a potential combination of prognostic biomarkers in pancreatic ductal adenocarcinoma using integrated bioinformatics analysis. PeerJ. 2020;8:e10419. doi: 10.7717/peerj.10419. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Yiqi Z, Ziyun L, Qin F, et al. Identification of 9-gene epithelial-mesenchymal transition related signature of osteosarcoma by integrating multi cohorts. Technol Cancer Res Treat. 2020;19:1533033820980769. doi: 10.1177/1533033820980769. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Zhang C, Liang Y, Ma MH, et al. KRT15, INHBA, MATN3, and AGT are aberrantly methylated and differentially expressed in gastric cancer and associated with prognosis. Pathol Res Pract. 2019;215(5):893–899. doi: 10.1016/j.prp.2019.01.034. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Material
IANN_A_2426758_SM3325.zip (10.3MB, zip)

Data Availability Statement

The data that support the findings of this study are available from the corresponding author, Jingjing Jing, upon reasonable request.

Articles from Annals of Medicine are provided here courtesy of Taylor & Francis

RESOURCES