Construction and Validation of a Ferroptosis-related Prognostic Signature in Kidney Renal Clear Cell Carcinoma

Background: Kidney renal clear cell carcinoma (KIRC) is the most common and lethal renal cell carcinoma (RCC) histological subtype. Ferroptosis is a newly discovered programmed cell death and serves an essential role in tumor occurrence and development. The purpose of this study is to analyze ferroptosis-related gene (FRG) expression proles and to construct a multi-gene signature for predicting the prognosis of KIRC patients. Methods: RNA-sequencing data and clinicopathological data of KIRC patients were downloaded from The Cancer Genome Atlas (TCGA). Differentially expressed FRGs between KIRC and normal tissues were identied using ‘limma’ package in R. GO and KEGG enrichment analyses were conducted to elucidate the biological functions and pathways of differentially expressed FRGs. Consensus clustering was used to investigate the relationship between the expression of FRGs and clinical phenotypes. Univariate and the least absolute shrinkage and selection operator (LASSO) Cox regression analysis were used to screen genes related to prognosis and construct the optimal signature. Then, a nomogram was established to predict individual survival probability by combining clinical features and prognostic signature. Results: A total of 19 differentially expressed FRGs were identied. Consensus clustering identied two clusters of KIRC patients with distinguished prognostic. Functional analysis revealed that metabolism-related pathways were enriched, especially lipid metabolism. A 7-gene ferroptosis-related prognostic signature was constructed to stratify the TCGA training cohort into high- and low-risk groups where the prognosis was signicantly worse in the high-risk group. The signature was identied as an independent prognostic indicator for KIRC. These ndings were validated in the testing cohort, the entire cohort, and the International Cancer Genome Consortium (ICGC) cohort. We further demonstrated that the signature-based risk score was highly associated with the KIRC progression. Further stratied survival analysis showed that the high-risk group had a signicantly lower overall survival (OS) rate than those in the low-risk group. Moreover, we constructed a nomogram that had a strong ability to forecast the OS of the KIRC patients. risk scores in individuals with more advanced stage disease. stratied survival the high-risk group had a signicantly lower OS rate than those in the low-risk group for all hierarchical cohorts. In addition, the signature of the seven FRGs was independent of other clinical factors. also provided a nomogram that reduced the ferroptosis-related signature combined with other clinical parameters into a single numerical estimate of the probability of an event to predict the prognosis of every individual patient. Taken together, these results signicantly indicated the potential role of ferroptosis in KIRC.

from traditional apoptosis or autophagic cell death or necrosis [13]. Ferroptosis is closely related to the metabolism of iron, fatty acids, amino acids, as well as the biosynthesis of glutathione, phospholipids, and NADPH [14,15]. Given that the irondependent accumulation of lipid reactive oxygen species (ROS) kills cells undergoing ferroptosis, both iron metabolism and lipid peroxidation are two critical processes involved in the mechanism of ferroptosis [16]. Preliminary evidence suggests that may ferroptosis may have a tumor suppressor function that could be potentially bene cial for cancer therapy [14]. On the other hand, various studies have also con rmed the pivotal role of ferroptosis in tumor development [17][18][19]. For example, Guerriero et al. evaluated the GPX4 expression in hepatocellular carcinoma tissue samples and veri ed that GPX4 was signi cantly over-expression and associated with an increased malignancy grade [20]. There is longstanding evidence that various primary tumors and also metastases express DPP4 to a variable extent [21]. In addition, other ferroptosis regulatory genes such as S1R [22], NRF2 [23], and HSPB1 [24] have also been shown to be strongly correlated with the tumorigenesis and progression. However, whether these ferroptosis-related genes (FRGs) are correlated with the prognosis of KIRC patients has not been investigated in detail.
In the present study, we rst performed a genome-wide comparative analysis of FRGs expression pro les and investigated differentiated FRGs expression patterns in KIRC patients based on the data obtained from The Cancer Genome Atlas (TCGA). Subsequently, we conducted functional enrichment analysis for these aberrantly expressed FRGs to systematically explored their potential functions and molecular mechanisms. After Cox univariate analysis and least absolute shrinkage and selection operator (LASSO) Cox regression analysis, we constructed a 7-gene signature with the ability to predict the prognosis of patients with KIRC in the training cohort and validated its prognostic value in the testing cohort, the entire cohort, and the ICGC cohort. Univariate and multivariate survival analysis indicated that the signature could serve as an independent factor for KIRC patients. Moreover, we developed a novel promising prognostic nomogram model with more accurate predictive ability by combining the ferroptosis signature and clinical parameters. Overall, our results demonstrate that some FRGs play vital roles in KIRC progression and are potential prognostic markers and therapeutic targets for KIRC.

Data acquisition
We downloaded the RNA-sequencing dataset of 72 normal kidney tissue samples and 539 KIPC samples with corresponding clinical data from The Cancer Genome Atlas (TCGA, https://cancergenome.nih.gov/) database. We excluded cases (n = 5) without follow-up records (survival time code of 0 months). Patients (n = 48) with incomplete clinical data were excluded from this analysis. RNA-seq data and clinical information of another 90 KIRC were obtained from The International Cancer Genome Consortium (ICGC, https://dcc.icgc.org/) data portal, which were used as an independent external validation set. Ethics approval and informed consent were not required for this study because the data from TCGA and ICGC were both publicly available. In addition, our research followed the access rules and publication guidelines provided by the TCGA and ICGC.
We searched the previous literature to identify 61 ferroptosis-related genes (FRGs) described so far to be involved in ferroptosis [14,19,25], and the gene sets of the ferroptosis process were shown in Additional le 1: Table S1. The expression data of the FRGs were extracted and used for subsequent analysis.
Data preprocessing and differentially expressed FRGs screening in KIRC All the raw expression data were subject to quality control, background correction, normalization, and logarithmic conversion by the R language (version 4.0.1, http://www.r-project.org/) [26] or Practical Extraction and Reporting Language (PERL, version 5.30.2, http://www.perl.org) [27]. The ensemble gene IDs were then converted to gene symbols using the GRCh38 reference genome (http://asia.ensembl.org/index.html) in this study. The FRGs that were differentially expressed between the tumor and normal samples were then identi ed based on a threshold of |logFC (fold change)| > 1 and a false discovery rate (FDR) < 0.05. Next, a Protein-Protein Interaction (PPI) network of the differentially expressed FRGs was generated by the Search Tool for the Retrieval of Interacting Genes (STRING, http://string.embl.de/) [28]. Visualization was then rendered using Cytoscape (http://www.cytoscape.org/) [29].

Functional enrichment analyses
The biological functions of these differently expressed FRGs were comprehensively detected by gene ontology (GO) enrichment, which comprised three terms: biological process (BP), cellular component (CC), and molecular function(MF).
The Kyoto Encyclopedia of Genes and Genomes database (KEGG) was applied to explore the signi cant pathways of differentially expressed FRGs. All enrichment analyses were carried out by utilizing the 'ClusterPro ler' R package [30]. Both P and FDR values < 0.05 were considered to be statistically signi cant.
GSCALite (http://bioinfo.life.hust.edu.cn/web/GSCALite/) database was a web-based analysis platform for gene set cancer analysis, which can help cancer research community to discovery cancer pathways and drugs [31]. In the current research, we also used the GSCALite database to analyze degree of the genes activation or inhibition of the classical pathway.

Non-negative matrix factorization consensus clustering
To investigate the relationship between the expression of FRGs and clinical phenotypes in KIRC, we clustered KIRC cohort from TCGA into different clusters by consensus expression of FRGs with 'ConsensusClusterPlus' in R [32]. Principal component analysis (PCA) was carried out to evaluate the gene expression patterns in the different subgroups. We also compared the OS difference between different clus ters by the Kaplan-Meier meth od in R. Chi-square test was used to compare the frequency distribution of age, gender, grade, AJCC (The American Joint Committee on Cancer) stages, TNM (Tumor, Node, Metastasis) stages between different clusters. N stage was not analyzed because of a large proportion of missing data.
Construction and validation of the ferroptosis-related prognostic risk signature The expression data of the differentially expressed FRGs were explored using the univariate Cox regression analysis to screen ferroptosis-related genes with prognostic values (P < 0.05). The TCGA-KIRC dataset was then randomly divided into a training cohort (n = 294) and a testing cohort (n = 192) for subsequent validation. To control the complexity of the model and avoid over tting, the least absolute shrinkage and selection operator (LASSO) Cox regression was performed to construct a prognostic signature within the training cohort [33]. We repeated the simulations 1000 times for which the optimal penalty parameter (λ) was selected through 10-fold cross validation following the minimum criteria. In addition, only genes with non-zero coe cients in the LASSO Cox regression were chosen to further calculate the risk score [34]. The risk score was estimated using the following formula: Risk Score. N, Expi, and Coei represented gene number, level of gene expression, and coe cient value, respectively. The median risk score was chosen as a cutoff value to dichotomize the training cohort into high-risk and low-risk groups. The Kaplan-Meier survival curve was plotted to evaluate the differences in OS between the two groups by the log-rank test. Additionally, the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC value) were used to evaluate the accuracy of the prognostic signature. Univariate and multivariate Cox regression analyses were used to deter mine whether risk score was an independent prognostic factor. For the survival analysis of each gene in the signature, the optimal cut-off expression value was determined using the X-Tile software [35]. Besides, the testing cohort, the entire cohort, and the ICGC cohort were used as validation cohorts to con rm the predictive capability of this prognostic signature using the same formula.
The clinical utility of the prognostic risk signature To test the usability and feasibility of the signature in the clinic, the associations between the ferroptosis-related risk signature and clinicopathologic features in TCGA-KIRC patients were performed. Using the Kaplan-Meier survival curves, we explored the ability of the prognostic risk signature to predict the survival outcomes of KIRC patients strati ed by various clinicopathological characteristics, including age (≤ 60 and > 60), gender (female and male), grade(G1-2 and G3-4), AJCC stage( / and / ), T stage (T1-2 and T3-4), and M stage (M0 and M1).
In addition to this, a nomogram, integrated the prognostic signature and clinical parameters, was constructed as a quantitative prediction tool to evaluate clinical prognosis. Following that, calibration curves were generated to evaluate the concordance between actual and predicted survival. Moreover, the concordance index (C-index) was computed to assess the model performance for predicting prognosis, with a C-index of 1 indicating perfect discrimination and a C-index 0.5 indicating a random guess. Decision curve analysis (DCA) was performed by calculating the net bene ts for a range of threshold probabilities to estimate the clinical usefulness of the nomogram.
The expression patterns, SNVs, CNVs, and drug sensitivity of the genes in the signature In order to con rm the reliability of the genes in the signature, we veri ed the expression patterns in different pathological tumors based on the data from the TCGA database. GSCALite database was used to analyze single nucleotide variations (SNVs) and copy number variations (CNVs) of the genes in the signature in KIRC patients.
CellMiner (https://discover.nci.nih.gov/cellminer/home.do) was used to detect the levels of the signature-related genes in the different renal cancer cells and the resulting values were represented as a heatmap. To provide support for drug selection of gene targeting therapy, we then analyzed the correlation of gene expression and drug sensitivity in KIRC patients with GSCALite database.

Differentially expressed FRGs in KIRC
Using transcriptome data from the TCGA database, we analyzed the mRNA levels of 61 FRGs in KIRC and normal tissues.
A total of 19 differentially expressed FRGs were eventually identi ed using the criteria of |log2FC| > 1 and FDR < 0.05, including 13 downregulated genes (MT1G, ACSF2, CHAC1, ACSL4, AKR1C2, PEBP1, PTGS2, AKR1C1, CBS, GOT1, ACO1, FDFT1, and HMGCR) and 6 upregulated genes (ALOX12, CD44, SLC7A11, ALOX5, HMOX1, and ALOX15B) in KIRC tissues compared to the normal kidney tissues (Fig. 1a, b). The box diagram was utilized to exhibit the expression patterns, median values, and data ranges of the differentially expressed FRGs in tumor and normal samples (Fig. 1c). The interaction network among these genes was presented in Fig. 1d, and the result indicated that PTGS2 and HMOX1 seemed to be the hub genes in this network.

Two ferroptosis subgroups were different in clinical phenotypes and OS via consensus clustering analysis
According to expression levels of 61 FRGs, the total TCGA-KIRC cohort was clustered into 2 subgroups (cluster 1 and cluster 2) with k = 2 as the optimal value because the grouping was suboptimal when they were divided into more than 2 clusters (Fig. 2a-c). Moreover, principal component analysis (PCA) was performed to compare the transcriptional pro le between cluster 1 and cluster 2. The result demonstrated that there was a signi cant distinction between the two subgroups ( Fig. 2d). To better understand the clustering result and its relationships with survival outcomes, we compared the OS between cluster 1 and cluster 2 and observed cluster 2 had shorter OS than cluster 1 for the KIRC patients (Fig. 2e).
We then evaluated associations between the clustering and the clinicopathological parameters of KIRC patients. The result showed that these 2 clusters were different in grade (P < 0.05), AJCC stage (P < 0.001), T stage (P <0.01), N stage (P <0.01), and survival status (P <0.05), but did not show any signi cant differences in age and gender (Fig. 2f). Therefore, these results suggested that ferroptosis was closely related to clinical phenotypes and the progression of KIRC.

Functional enrichment analyses of the differentially expressed FRGs
To elucidate the biological functions and pathways of the 19 differentially expressed FRGs, GO functional annotation and KEGG pathway enrichment analyses were conducted. The results indicated that these differently expressed FRGs were signi cantly enriched in the biological process (BP) related to several metabolic process, such as cofactor metabolic process, fatty acid metabolic process, and fatty acid derivative metabolic process (Fig. 3a, b). Furthermore, lipoxygenase pathway was also involved. In terms of molecular function (CC), we found that the differently expressed FRGs were signi cantly enriched in peroxisomal membrane, microbody membrane, and caveola. Through the molecular function (MF), the differently expressed FRGs were notably enriched in oxidoreductase activity, dioxygenase activity, and lyase activity. In the KEGG pathway enrichment analysis, these genes were shown to be mostly associated with pathways in arachidonic acid metabolism, serotonergic synapse, and ferroptosis (Fig. 3c, d). In addition, we explored the effect of the differentially expressed FRGs in multiple classical signaling pathways on KIRC using the GSLA database. The results revealed that some FRGs was associated with the activation or inhibition of oncogenic pathways. For example, SLC7A11 expression was related to the activation of apoptotic, cell cycle, and EMT pathways; PTGS2 expression was related to the inhibition of cell cycle, DNA damage response, hormone AR, and hormone ER (Additional le 2: Figure S1).
Prognosis-related FRDs selecting and construction of a prognostic signature based on seven FRGs in the TCGA training cohort of KIRC patients We conducted a univariate Cox regression analysis on the transcriptome data from the TCGA-KIRC dataset and found that the expression of 11 FRGs (CBS, GOT1, FDFT1, HMOX1, CD44, ACO1, AKR1C2, PEBP1, CHAC1, HMGCR, and SLC7A11) was signi cantly related to the KIRC survival (P < 0.05; Fig. 4a). Genes (CBS, CD44, AKR1C2, CHAC1, and SLC7A11) with HR > 1 were considered as risk genes, while the remaining six genes (GOT1, FDFT1, HMOX1, ACO1, PEBP1, and HMGCR) with HR < 1 as protective genes. Based on the expression pro le of the 11 genes mentioned above in the training cohort, a prognostic risk signature was constructed by the least absolute shrinkage and selection operator (LASSO) Cox regression analysis. As a result, a 7-gene signature (CBS, HMOX1, CD44, AKR1C2, CHAC1, HMGCR, and SLC7A11) was identi ed based on the optimal value of λ (Fig. 4b, 4c). Survival analyses based on the optimal cut-off expression value of each gene showed that high expression of risk genes (CBS, CD44, AKR1C2, CHAC1, and SLC7A11) was correlated with poor prognosis, while high expression of protective genes (HMOX1 and HMGCR) displayed the opposite patterns (Additional le 3: Figure S2).
Based on the 7 candidate FRGs, the risk score of each patient was calculated according to the following formula: Risk score = (3.8463 × CBS) + (-0.0021 × HMOX1 ) + (0.0101 × CD44) + (0.0208 × AKR1C2) + (0.0252 × CHAC1) + (-0.1354 × HMGCR) + (0.3774 × SLC7A11). We then used the median risk score as a cutoff point for classifying KIRC patients in the training cohort (n = 294) into high-risk group (n = 147) and low-risk group (n = 147). Kaplan-Meier survival curve analysis showed that OS was signi cantly different between the predicted two risk groups and the high-risk group had signi cantly shorter survival time compared to low-risk group (P = 4.709e-06; Fig. 4d). The survival rates at three and ve years in the high-risk group were 61.9% and 49.2%, respectively, while the corresponding rates in the low-risk group were 86.3% and 76.7%, respectively. The ROC curve analysis indicated the risk signature had a promising predictive value for KIRC survival prediction (AUC = 0.734) (Fig. 4e). We then ranked the risk scores of the patients and then analyzed their distributions (Fig.   4f). The distributions of risk scores and the survival status suggested survival rate and time were signi cantly increased in the low-risk group compared to the high-risk group (Fig. 4g). The heatmap displayed that the expression levels of the 7 candidate FRGs in the high-risk and low-risk groups (Fig. 4h).
Validation of the prognostic signature in the TCGA testing cohort, the TCGA entire cohort, and the ICGC cohort To verify the accuracy and robustness of the prognostic risk signature, the prognostic value of the signature was further validated in the testing cohort (n = 192), the entire TCGA-KIRC cohort (n = 486). The prognostic risk score was calculated for patients in each cohort based on the prognostic signature. Patients of the testing cohort (n = 192) were strati ed into the high-risk group (n = 96) and low-risk group (n = 96) with the median risk score as the cutoff. Similar to the procedure in the entire cohort, the 489 patients of the entire cohort were divided into the high-risk group (n = 243) and low-risk group (n = 243). The detailed clinical features of all KIRC patients were listed in Table 1.
We observed that the results in the testing and entire cohort were consistent with the outcome in the training cohort.
Kaplan-Meier survival curves showed that patients in the high-risk group had a signi cantly worse OS than their low-risk counterparts in both the testing cohort (P = 2.149e-08) (Fig. 5a) and the entire TCGA-KIRC cohort (P = 7.724e-13) (Fig. 5d). In the testing cohort, the survival rate at 3 years in the high-risk and low-risk group was 65.1% and 87.9%, respectively. The survival rate of high-risk group was 32.4% at 5 years, and that of the low-risk group was 84.8%. Similarly, in the entire TCGA-KIRC cohort, the survival rates at 3 and 5 years in high-risk group were lower than those in the low-risk group (63.4 vs. 87.7% at 3 years, 41.3 vs. 80.2% at 5 years). ROC curve analysis showed that the AUC values for in the testing cohort and the entire TCGA-KIRC cohort were 0.762 (Fig. 5b) and 0.749 (Fig. 5e), respectively. The risk score distribution, survival status, and the risk gene expression in the testing cohort and the entire TCGA-KIRC cohort were shown in Fig. 5c and 5f.
Then the prognostic signature was validated in the ICGC cohort (n = 90). Patients were divided into high-risk group (n = 45) and low-risk group (n = 45) with the median risk score. The OS was signi cantly poorer in the high-risk group than in the low-risk group (P = 1.592e-02) (Additional le 4: Figure S3A). The AUC value for the prognostic signature was 0.71, suggesting well-prediction performances (Additional le 4: Figure S3B). The distribution of risk score, survival status, and gene expression of KIRC patients in the ICGC cohort was presented in Additional le 4: Figure S3C, which were similar to the above cohorts. Taken together, these results revealed that the risk signature could accurately predict the prognosis of KIPP patients.

The prognostic signature is an independent prognostic factor for KIRC patients
To determine whether the risk signature can be used as an independent prognostic factor, univariate and multivariate Cox regression analyses were performed with the risk score and relevant clinical factors, including age, gender, grade, AJCC stage, T stage, and M stage. In the TCGA training cohort, univariate analyses showed that age and risk score were signi cantly associated with OS. Besides, subsequent multivariate analyses suggested that age and risk score were still signi cantly associated with OS (Table 2). Similar results were obtained for both the testing and the entire TGCA-KIRC cohorts (Table 2). Therefore, the risk score calculated based on the prognostic risk signature was an independent adverse prognostic factor for OS in KIRC patients.

Prognostic risk score indicated strong associations with clinical characteristics in KIRC
To investigate whether the risk signature could better predict KIPC clinicopathological features, an analysis was performed to explore the associations between the risk signature and clinical parameters. Signi cant differences were observed between two groups in grade (P = 6.264e-06) (Fig. 6a), AJCC stage (P = 1.6973-06) (Fig. 6b), T stage (P = 4.884e-06) (Fig.  6c), and M stage (P = 0.002) (Fig. 6d). Simultaneously, we observed that advanced-stage tumors were signi cantly associated with the high-risk group, whereas, the early-stage tumors were correlated with the low-risk group.
To detect the prognostic value of our risk score model in different subgroups, we further investigated strati ed survival analysis using the following clinical variables: age (≤60 and >60), gender (female and male), tumor grade (G1-2 and G3-4), AJCC stage (I & II and III & IV), T stage (T1-2 and T3-4), and M stage (M0 and M1). Interestingly, survival analysis indicated the high-risk group had a signi cantly lower OS rate than those in the low-risk group for all hierarchical cohorts (Fig. 7). Thus, these results suggested that the classi cation of the risk signature can be used to precisely identify patients with poor prognosis, without the consideration of clinical parameters.

Development of a personalized prognostic nomogram
A nomogram is a powerful tool that has been extensively applied to quantitatively determine individuals' risk in clinical decision making by incorporating multiple clinical factors [36,37]. To establish a viable method for predicting survival in patients with KIRC, we developed a prognostic nomogram based on the constructed prognostic risk signature and serval clinical features, including age, gender, grade, AJCC stage, T stage, M stage. The nomogram was devoted to evaluate the probability of 1-, 3-, and 5-year survival (Fig. 8a). Each factor was assigned a score in proportion to its risk contribution to survival in the nomogram. The C-index that was used to evaluate the OS of the nomogram was 0.771. Calibration curves showed optimal agreement when compared with an ideal model (Fig. 8b), particularly for 3-and 5-year survival predicted probabilities. Decision curve analysis (DCA) of the nomogram indicated that the nomogram had a wide and practical range of threshold probability for the TCGA-KIRC cohort for predicting survival rates (Fig. 8c).
The expression patterns, SNVs, CNVs of the seven candidate genes in the signature for KIRC patients Subsequently, we analyzed the correlation between the expression of each of the seven prognostic risk signature genes and the clinicopathological features of KIRC patients in the TCGA cohort. In terms of grade alone, CBS, CD44, and CHAC1 increased with tumor grade, while HMGCR was decreased. No signi cantly difference in the expression values of HMOX1, AKR1C2, and SLC7A11 was detected between different tumor grades (Fig. 9a). As for different AJCC stage, CBS, HMGCR, CHAC1, and HMOX1 were signi cant differentially expressed, with higher expression levels of CBS and CHAC1 indicating higher advanced AJCC stage, while HMGCR and HMOX1 showed the opposite trend (Fig. 9b). Regarding T stage, it was noted that CBS, CD44, CHAC1, and SLC7A1 were signi cantly up-regulated in advanced T grade, whereas HMGCR was signi cantly down-regulated (Fig. 9c). CBS, CD44, and HMGCR also represented similar trends in N stage as T stage (Fig.  9d). In general, the expression levels of CBS, CD44, CHAC1, and SLC7A1 were positively correlated with tumor progression, while HMOX1 and HMGCR were negatively correlated with tumor progression, which was consistent with the above study.
In addition, AKR1C2 expression appeared to be independent of tumor progression.
We then used the GSCA database to determine the single nucleotide variations (SNVs) of the seven candidate genes in the signature for KIRC patients. The results indicated that the most frequent mutation type was single nucleotide polymorphism (SNP) (Additional le 5: Figure S4A), and missense mutation was the most fraction among these mutations (Additional le 5: Figure S4B). In addition, C > T transversion accounted for the most common of SNV (Additional le 5: Figure S4C). The characteristic of the frequently mutated genes was showed in Additional le 4: Figure S4D. We also analyzed the copy number variations (CNVs) of the seven candidate genes in the signature and found heterozygous mutations (ampli cation and deletion) in all genes (Additional le 5: Figure S4E). Drug sensitivity of the seven candidate genes in the signature for KIRC Patients We next utilized the CellMiner database to explore the expression of seven candidate genes in diverse kidney cancer cell lines, including 786-O, A498, ACHN, CAKI-1, RXF 393, SN12C, UO-31, and TK-10. We observed that the expression levels of these genes showed great heterogeneity in different cell lines (Additional le 6: Figure S5A). In addition, we also investigated the drug sensitivity of the seven candidate genes for the KIRC patients using the GSCALite database. Four of them (CD44, SLC7A11, AKR1C2, and HMOX1) were highly related to drug sensitivity to a number of chemotherapy drugs (Additional le 6: Figure S5B), which provided direct support for drug targeted therapy.

Discussion
It has been previously reported that KIRC is a malignant disease, the pathogenesis may be related to the reprogramming of energetic metabolisms, such as tricarboxylic acid cycle, aerobic glycolysis, amino acids, fatty acids, and dysfunctional oxidative phosphorylation [38,39]. Ferroptosis is a programmed cell death caused by iron-dependent lipid peroxidation and is different from other types of cell death, including apoptosis [14,40]. A wide variety of human diseases, including cancer, have been associated with the abnormal function of ferroptosis, and inhibition or upregulation of ferroptosis modulates the metabolic reprogramming of cancer cells [14,41]. At present, few studies have been performed on ferroptosis in KIRC, and the results also remain controversial [42]. Therefore, investigating the expression patterns of FRGs is critical to understand the role of ferroptosis in KIRC.
In the current study, we systematically investigated the RNA-seq-expression and clinical information of KIRC from the TCGA database. We found that 19 out of 61 FRGs were differentially expressed in KIRC patients, including 6 up-regulated and 13 down-regulated genes. To gain more insights into the functional roles of the differentially expressed FRGs in KIRC, we performed functional enrichment analysis to investigate the associated biological processes and pathways. Functional annotation showed that many biological processes and pathways related to metabolism were enriched, especially lipid metabolism. The current consensus is that the execution of ferroptosis could result from the direct effects of lipid peroxidation [14]. Moreover, the increased lipid peroxidation is a principal mechanistic pathway in renal carcinogenesis induced by different chemicals. Therefore, we have reason to believe that ferroptosis may have a close connection with tumor metabolic reprogramming. We then used consensus clustering analysis to categorize the KIRC cohort into two subgroups (cluster1 and cluster 2) based on the expression of FRGs. Interestingly, the OS was dramatically different between the two subgroups, indicating that the levels of FRGs were closely associated with the prognosis of KIRC patients.
Using Lasso regression and Cox survival analyses, we constructed a risk prognostic signature based on seven prognostic FRGs (CBS, HMOX1, CD44, AKR1C2, CHAC1, HMGCR, and SLC7A11). Every patient with KIRC was assigned into high-risk and low-risk groups according to the prognostic risk signature. We demonstrated that the OS was shorter for the high-risk patients in four cohorts (the training cohort, the testing cohort, the entire TCGA-KIRC cohort, and the ICGC cohort) compared to the low-risk patients. The ROC curves and AUCs indicated that the signature performed well. Recently, more and more studies have suggested that the abnormal expression of the FRGs is involved in human cancer [16,17,41]. We further demonstrated that the signature-based risk score was highly associated with the KIRC progression. It was also observed that the risk scores were higher in individuals with more advanced stage disease. Further strati ed survival analysis showed that the high-risk group had a signi cantly lower OS rate than those in the low-risk group for all hierarchical cohorts. In addition, the signature of the seven FRGs was independent of other clinical factors. We also provided a nomogram that reduced the ferroptosis-related signature combined with other clinical parameters into a single numerical estimate of the probability of an event to predict the prognosis of every individual patient. Taken together, these results signi cantly indicated the potential role of ferroptosis in KIRC.
Our results showed that the expression level of CBS, CD44, AKR1C2, CHAC1, and SLC7A11was positively associated with the progression of KIRC, while contrary results appeared in HMOX1 and HMGCR. Cystathionine β-synthase (CBS), a fundamental enzyme in L-cystathionine synthesis, catalyzes the condensation of serine and homocysteine to form cystathionine, which is the initial and rate-limiting step in the transsulfuration pathway [43]. An increasing body of evidence points to the key roles of CBS in tumor progression, such as prostate cancer [44], ovarian cancer [45], and colon cancer [46]. However, another research strongly supports a negative regulatory role for CBS in hepatocellular carcinoma [47]. Thus, the biological function of CBS in cancer is complex and requires further investigation. CD44 is an important cancer stem cell marker in tumors and implicates in malignant processes including cell motility, tumor growth, and angiogenesis [48]. In fact, CD44 has been observed in many human tumors and associated with a poor survival rate [49]. Aldo-Keto reductase 1C2 (AKR1C2), a member of Aldo-Keto reductase subfamily, could mediate similar prostaglandin D2 conversion toward the accumulation of proliferative signals through PI3K/Akt signaling pathway to promote prostate cell proliferation [50]. In addition, Zhang. et al demonstrated that AKR1C2 could act as a targetable oncogene in esophageal squamous cell carcinoma via activating PI3K/AKT signaling pathway. ChaC glutathione-speci c γ-glutamyl cyclotransferase 1 (CHAC1) is a proapoptotic γ-glutamyl cyclotransferase that depletes glutathione. There are few studies on CHAC1 at present, and its clinical signi cances and biological functions in tumors remain unknown. Solute carrier family 7 member 11 (SLC7A11; also known as xCT) is a cystine/glutamate antiporter that imports cystine into the cells while exporting glutamate [51]. SLC7A11 is highly expressed in human tumors, and its overexpression inhibits ROS-induced ferroptosis and abrogates p53(3KR)-mediated tumor growth suppression in xenograft models [52]. Recent studies demonstrated that smoking could induce the expression of SLC7A11 in oral cancer cells, suggesting that overexpression of SLC7A11 may support lung tumor progression [53]. These results suggest that the four genes (CBS, CD44, AKR1C2, and SLC7A11) above are closely related to the occurrence and development of tumors, which are partly consistent with our research.
Heme oxygenase-1 (HMOX-1), a phase II enzyme that responds to electrophilic stimuli, has been reported to play protective or detrimental effects in different diseases, including cancers. HMOX-1 is elevated in a variety of human malignancies, indicating that it contributes to settle the tumor microenvironment for cancer cell growth, angiogenesis, and metastasis [54]. However, emerging evidence has revealed HMOX-1 functions as a negative regulator in erastin-and sorafenib-induced hepatocellular carcinoma and knockdown of HMOX-1 by speci c shRNA increased erastin-and sorafenib-induced growth inhibition [23]. Hydroxymethylglutaryl-coenzyme A reductase (HMGCR), the rate-limiting enzyme in the mevalonate pathway, which is generally believed to be a candidate metabolic oncogene [55]. For example, Li. et al found that HMGCR is up-regulated in gastric cancer and promotes the growth and migration of the cancer cells by activating Hedgehog/Gli1 signaling [56]. Interestingly, another study reported that oral atorvastatin and its metabolites are detectable in human breast samples, suggesting that HMGCR may be directly inhibited in breast tumors [57]. It is evident from the above study that HMOX-1 and HMGCR have both tumor-promoting and tumor-suppressive properties. Clearly, further research will be needed to elucidate the role of cancers, especially KIRC.
However, despite these encouraging results, there existed some limitations. Firstly, since the retrospective nature of this study, a multicenter and prospective study is required to assess the potential applicability of our conclusions. Secondly, further studies including in vivo and in vitro experiments are needed to elucidate the detailed mechanism of key genes for clinical applications. In addition, many prominent prognostic genes in KIRC might have been excluded because merely a single phenotype was considered to construct a prognostic signature in our study.

Conclusion
In conclusion, we provided insights into the roles of FRGs in KIRC and constructed a promising risk prognostic signature that exhibited potential as a biomarker of OS in KIRC patients. The underlying mechanisms whereby differentially expressed FRGs exert their biological roles in metabolism-associated biological processes. We then established a novel promising prognostic nomogram incorporating for providing individualized survival prediction. Therefore, our constructed ferroptosis-related signature is of great clinical importance and may help facilitate personalized medicine in the clinical setting.