The flow of this study is shown in Figure 1. We downloaded what added up to 407 GC patients from the TCGA database, including 375 GC samples and 32 normal para-cancer samples. The detailed clinical features of these patients are shown in Table 1.
Acquisition of prognostic ferroptosis-related DEGs
What added up to 259 ferroptosis-related genes were extracted from the FerrDb database(Figure 2a). Among them, 170 genes showed significant expression differences between GC samples and normal para-cancer samples, and 18 prognostic genes were screened out by univariate COX regression analysis (FDR<0.05, Figure 2). The heat map and forest map of the 18 genes are shown in figures 2b-c. The PPI network of these 18 genes was constructed in the STRING database, and it was found that CAV1, DUSP1, SP1, and NOX4 were the critical genes in this PPI network (Figure 2d). In addition, the correlation network of the 18 genes is indicated done Figure 2e.
Construction and validation of the prognostic model
Those 18 prognostic ferroptosis-related genes were further screened by LASSO regression analysis. The optimal λ value was calculated, and the corresponding 11 genes with the most reduced cross-validation lapse were chosen (figures 3a-b). Finally, a model was constructed based on the 11 genes. The risk score has been computed dependent upon each gene's expression level and the corresponding correlation coefficient (coef) (Table 2), and the patients were partitioned under a high-risk group and a low-risk group with the risk values' median value. (Figure 4a). The survival prognosis of the low-risk group is higher than that of the high-risk group through the Kaplan-Meier survival curve analysis(p=1.651e-04) (Figure 4b). The area under the AUC curve in the ROC curve was used to evaluate the model's prediction reliability, and as shown in Figure 4c, the model had good sensitivity. PCA and t-SNE analyses showed that patients' distribution in the two risk groups was in two directions (Figure 4d-e). Besides, patients with high-risk died earlier than the low-risk group (Figure 4f).
Analysis of the independent prognostic value of the model
To determine whether the established prognosis model can be used as a prognostic factor independent of other clinical features, univariate and multivariate COX regression analysis was used for verification. The results showed a significant correlation between risk scores and OS in both univariate and multivariate Cox regression analyses (HR=3.627, 95% CI=2.256-5.829, p<0.001, Figure 5a; HR=4.176, 95% CI=2.588-6.736, P<0.001, Figure 5b). Previously summary, the risk score of the prognostic model we constructed can be used as an independent prognostic factor.
Function analysis of the prognosis model
GO and KEGG analyses of DEGs between the two risk groups were performed to dissect related biological functions and pathways. Those outcomes indicated that the DEGs were significantly enriched in extracellular matrix (ECM) related biological processes, including ECM organization, collagen-containing ECM, focal adhesion, ECM structural constituent, collagen binding, ECM binding (P. adjust<0.05, Figure 6a). KEGG pathway analysis showed that these DEGs were also enriched in ECM related pathways, such as proteoglycans in cancer and ECM-receptor interactions (P. adjust<0.05, Figure 6b). Besides, these DEGs are enriched in the muscle system.
Tumor-targeted immunotherapy is a hot topic nowadays. ECM has been speculated as a part of health and homeostasis, plays a critical immunomodulatory role, and the synergistic antitumor effect between ECM and the immune system will be the focus of future research.
Therefore, ssGSEA is used to quantify the enrichment scores of different immune cell subsets, related functions, or pathways further to explore the correlation between risk scores and immune status. The results showed significant differences in mast cells, neutrophils, T helper cells, and Th2 cells between these two different risk groups (P. adjust<0.05, Figure 7a). In the KEGG analysis, except for the fact that MHC class I had a lower score in the low-risk group, others including APC co-stimulation, cytokine-cytokine receptor interaction (CCR), typeⅠIFN response, parainflammation, and type Ⅱ IFN response were all scored higher in the high-risk group (P. adjust<0.05, Figure 7b).