Identification of SDG and patients’ clinical data
A total of 9 normal and 78 EAC samples with gene expression profiles and clinical information were retrieved from TCGA dataset (Additional files 4 and 5). After analysis, there were 28 significantly different ferroptosis-related genes between normal and EAC samples. Among these, four FRG (AKR1C1, AKR1C2, MT1G and NFE2L2) were down-regulated in EAC tissues compared with normal tissue, other 24 genes were up-regulated (Table 1). The heatmap and deviation plots are shown in Fig. 1a, b.
Functional enrichment analysis of SDG
To elucidate the biological functions and pathways of SDG in ferroptosis, the 28 genes were used to functional enrichment analysis. The GO results showed that SDG were enriched in iron-related pathways, such as metabolic and oxidative process. KEGG analysis showed the SDG were closely enriched in ferroptosis, including the GSH metabolism, oxidative reaction and biosynthesis (Fig. 2a–d).
Interactions and correlations of SDG
We explored the SDG interaction at the STRING online website, and the gene network demonstrated the TP53, G6PD, NFE2L2 and PTGS2 were the hub genes (Fig. 3a). The correlations between these SDG are presented in Fig. 3b.
Prognostic FRG and independent risk factors
Nearly half of the FRG (46.67%, 28/60) were differently expressed in the normal and EAC tissues. By performing the univariate cox regression analysis in the 78 EAC patients, we identified four FRG (CARS1, GCLM, GLS2 and EMC2) were significantly associated with OS (all P < 0.05) (Fig. 4a). Subsequently, multivariate cox regression analysis indicated GLS2 was independent prognostic risk factor (HR = 6.328, P = 0.004) (Fig. 4b).
According to the median of the risk score (risk score formula = 0.781 * expression level of CARS1 + 0.474 * expression level of GCLM + 1.845 * expression level of GLS2 + 0.717 * expression level of EMC2), patients were stratified into high and low-risk groups. Then, we explored the clinical information values in the patients’ OS combining with the FRG. In the univariate cox analysis, we found that tumor stage and risk score were significantly associated with OS (all P < 0.001) (Fig. 4c). And the multivariate cox regression showed the tumor stage and risk score were independent risk factors in EAC patients’ OS (HR = 6.755, HR = 1.328, all P < 0.001 respectively) (Fig. 4d).
Prognostic hazard curves in high and low-risk patients
Seventy-eight EAC patients were divided into high-risk group (n = 39) and low-risk patients (n = 39) according to the median of the risk score. The Kaplan–Meier curve shows patients with high-risk score have a significant higher death probability than those with low-risk (median time = 0.657 years vs. 1.192 years, p = 0.0075, Fig. 5a). As the risk score increases, the patients’ death risk increases, AND the survival time decreases (Fig. 5b, d). The risk heatmap clearly shows EMC2 was up-regulated in high-risk group compared with the low-risk group, implying it is a tumor-promoting role (Fig. 5c).
Construction of predictive models
In order to provide an applicable method to predict the EAC patients’ OS, we established the ROC curve using the independent risk factor (risk score) from the multivariate cox regression.
In addition, we also assessed the feasibility using the area under curve (AUC) value. The results showed the risk score had better predictive ability (AUC = 0.744) (Fig. 6).
Immune cell enrichment analysis
To further explore the relationships between the risk scores and immune cells and functions, we quantified the enrichment scores of 16 immune cell subpopulations and their related functions with the ssGSEA R package. The results showed the types of immune cells (such as DCs, iDCs, mast cells, Th2 cells, TIL cell, Treg cells, B cells, CD8+ T cells, pDCs, T helper cells, Th1 and Tfh cells) in the high-risk group were significant different with those in the low-risk group (Fig. 7a). Moreover, the scores of the immune functions, such as the type I IFN response, type II IFN response, T cell co-inhibition, APC inhibition and check-point were significantly higher in low-risk group, implying their immunological functions associated with ferroptosis were more active in the low-risk group (Fig. 7b).
To better understand characteristics of immune cells and their relations with FRG, the TIMER database was used to analyze the correlation between the abundance of immune cells and the four prognostic genes (CARS1, GCLM, GLS2 and EMC2). The results show GCLM is negatively associated with CD8+ T cell and neutrophil (P = 0.047, 0.020 respectively). GLS2 have negative correlations with CD8+ T cell and neutrophil (P = 0.025, 0.025 respectively), positive with B cell (P = 0.013). EMC2 (also known as TTC35) is negatively correlated with B cell, CD4+ T cell and neutrophil (P = 0.038, 0.006, 0.027 respectively). The details are shown in Fig. 8a–d.
Clinical experimental validation
We performed the PCR and IHC validation in clinical specimens following the steps described above. We verified the five most significantly different genes according the logFC values (NOX1, MT1G, PTGS2, ALOX5, TFRC). By analysis, the PCR results showed MT1G was down-regulated and the ALOX5 and NOX1 were significantly up-regulated in the EAC tissues. There were no significant differences in the expression of PTGS2 and TFRC between the normal and EAC specimens. The details of the five genes were visualized in Fig. 9a–e (Additional file 6).
Immunohistochemistry results showed ALOX5, NOX1, PTGS2 and TFRC proteins were expressed at high frequency in EAC tissues compared to normal tissues (Fig. 10a–h). Interestingly, the PTGS2 and TFRC protein levels were not consistent with the PCR results. This may be related to the gene transcription regulation and small sample size. The complete IHC results are provided in Additional file 7: Figure S1.