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

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; deep learning; non-apoptotic regulated cell death; prognosis; therapy.

Figure 1.

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).

Figure 2.

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).

Figure 3.

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.

Figure 4.

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).

Figure 5.

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).

Figure 6.

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.

Figure 7.

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).

Figure 8.

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).

Figure 9.

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.