A novel ferroptosis-related genes model for prognosis prediction of lung adenocarcinoma

LUAD is a frequently detected subtype of NSCLC. LUAD dataset obtained from the TCGA database (https://​portal.​gdc.​cancer.​gov/​) involved 533 cancer specimens and 59 paracancerous ones. Their raw RNA-Seq data, single-nucleotide variation (SNV) data and clinical information were included as well. The mRNA expression data were normalized using the DESeq2 variance stabilizing transformation (VST). Based on the previous research, we selected the top 60 ferroptosis-related genes as the candidate gene set (Additional file 5: Table S1) [4, 13‐15].

The DESeq2 package in R [16] was utilized to analyze DEGs between LUAD specimens and paracancerous ones based on the false discovery rate (FDR) < 0.05. The Cox proportional-hazards model is a type of semiparametric regression model, in which survival outcomes and survival time are considered as the dependent variables, and the impact of multiple factors on survival time is analyzed. The most crucial concept in the model is hazard ratio (HR). Generally speaking, HR > 1 and HR < 1 indicate a risk and protective prognostic factor in cancer dataset, respectively. The values in the matrix of VST were utilized to the following univariate Cox regression analysis that assessed potential influences of ferroptosis-related genes on the overall survival of LUAD. P values were adjusted with the Benjamini & Hochberg (BH) procedure. Genes with HR ≠ 1 and p < 0.05 were selected as prognosis-related genes. Subsequently, the intersection set between the identified DEGs and prognosis-related genes was taken, and their downstream functions in influencing the prognosis of LUAD were further analyzed.

The glmnet package in R was utilized to perform LASSO regression analysis on the target gene set [17] and the establishment of the multi-gene prognostic prediction model. To avoid over-fitting and obtain reliable results, we applied the tenfold cross-validation to acquire optimal lambda values from the minimum partial likelihood deviance, and screened out representative genes. A few candidates were selected to establish a multivariate Cox regression model [18]. Using the coxph function, the PH test of each factor was performed, and the VIF and correlation coefficient of each factor in each regression model were calculated as well. Meanwhile, the possible collinearity of factors was determined. Variables that conformed to the PH hypothesis and collinearity tests were re-modeled. Risk scores of LUAD patients were calculated according to the modeling results. The formula was established as follows: Risk score = Sum (expression level of each gene × corresponding coefficient). Expression level of each gene was the normalized mRNA expression, and the coefficient was the result of the multivariate Cox regression analysis.

To explore the survival predictive feasibility of risk scores obtained from the multi-gene prognostic prediction model, we used the "survival" and "survminer" packages and the Kaplan–Meier method to estimate the survival curve. The time-dependent ROC curve was plotted by the "survival ROC" package in G, which graphically displayed the predictive ability in the prognosis of LUAD at different time points.

To establish a more reliable prediction method that could be applied in clinical practice, we constructed the composite nomogram through the rms package. Similarly, clinical features and risk scores of LUAD patients were combined to establish a multivariate Cox regression model. Based on the influenced degrees of risk factors on the outcome, they were graded to value its level, and a total score was obtained by the sum of the score of each factor. Finally, through the converse relation between the total scores and the outcome events, the predicted survival time of LUAD patients was calculated.

We further verified the accuracy of the nomogram by applying the model using the C-statistic, calibration curve and time-dependent ROC curve. Meanwhile, the nomogram model was calibrated to predict the proximity between the predicted survival and the actual result numerically. The Hosmer–Lemeshow (H–L) test was performed by dividing data into 3 ascending ordered groups (tertiles) based on the predicted result obtained from the model, thus testing the goodness of fit for the χ2 test. Furthermore, the average predicted survival was compared with the actual event rate estimated using the Kaplan–Meier method.

DEGs between high-risk groups and low-risk groups divided by risk scores were obtained. Subsequently, enrichment analysis of gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways of DEGs were analyzed using the clusterprofiler package in R [19].

The single-sample gene set enrichment analysis (ssGSEA) on 29 immune gene sets [20] involving 16 immune cells and 13 immune-related activation pathways was performed using the gsva package in R [21] (Additional file 6: Table S2). The degree of immune infiltration of each sample was recorded.

The sample function in R was used to perform the random stratified sampling. Meanwhile, the heatmap of DEGs and SNV mutation landscape were respectively depicted by the pheatmap and maftools packages in R [22]. Wilcoxon-test was used to test the difference between groups of the boxplot. Kaplan–Meier survival curve was depicted for assessing the prognostic potential, followed by the log-rank test to compare the difference between curves. The R software (Version 3.6.0) was used for all statistical analyses. The flowchart of this study was shown in Fig. 1.

The downloaded LUAD dataset contained 533 LUAD specimens and 59 paracancerous ones. Among them, 502 LUAD cases had clinical information of the overall survival (Table 1). Besides, the DESeq2 package was used to analyze the mRNA expression data of the original counts for obtaining DEGs (FDR < 0.05). The acquired DEGs were sorted according to the value of log2 fold change, and they were displayed by depicting the heatmap (Fig. 2a). Univariate Cox regression analysis revealed that among the 60 ferroptosis-related genes, 11 were significantly correlated to the prognosis of LUAD (adj p < 0.05). Taking the intersection of the 11 genes and DEGs, it is obviously shown that the former ones were differentially expressed between LUAD and paracancerous specimens. In addition, the results of Cox regression analysis were depicted in the forest plots (Fig. 2b).

We randomly stratified LUAD patients according to their tumor stage and divided them into the training set (n = 251) and validation set (n = 251). Meanwhile, LASSO-Cox regression analysis of the above-mentioned set was performed to identify prognosis-related genes that were correlated to ferroptosis (Fig. 3a, b). The results revealed that a total of 5 genes conformed to the pH-partition hypothesis (Additional file 1: Figure S1) and the collinearity test (correlation coefficient less than 0.5) (Fig. 3c), namely the ACSL4, GSS, ACSL3, PEBP1 and PGD genes. Through plotting the forest plots of them, it is disclosed that the ACSL4 (adj P < 0.05), GSS, ACSL3 and PGD genes were the risk factors for the prognosis of LUAD, while the PEBP1 gene was a protective factor (adj P < 0.05) (Fig. 3d). Furthermore, LUAD patients were subgrouped based on the expression level of the 5 identified genes, and the corresponding Kaplan–Meier survival curves were displayed in Fig. 4. Besides, the threshold of the expression level of each gene was calculated by the surv_cutpoint function from the survminer R package (Additional file 2: Figure S2). We subsequently calculated the risk score of each sample based on the modeling results using the following formula: Risk score = 0.2226 × expression level of ACSL4 + 0.2373 × expression level of GSS + 0.4369 × expression level of ACSL3—0.4417 × expression level of PEBP1 + 0.1431 × expression level of PGD. Expression level of each gene was normalized. LUAD patients were further divided into the high-risk group and low-risk group based on the individualized risk score. The grouped threshold was determined by the median value of the risk score of all LUAD patients in the training set, the median of which was 6.64. Kaplan–Meier survival curves based on the overall survival calculated using the KM algorithm have indicated that LUAD patients in the high-risk group had a worse survival than that of the low-risk group (p = 0.0041). Later, the predictive potential of the risk score model in LUAD was assessed by depicting the ROC curve, and the highest AUC was 0.7 (Fig. 5a).

To determine the reliability of the established model and the predictive accuracy, results obtained from the prognostic model were validated in the verification set. Similarly, risk scores in the validation set, and subgroup classification of LUAD patients into the high-risk group and low-risk group with the same threshold (cut-off = 6.64) were conducted. Moreover, the survival status of LUAD patients in the validation set was estimated using the KM algorithm, followed by plotting the survival curve. It is consistently shown that the prognosis of LUAD patients in the high-risk group was significantly worse than that of the low-risk group (p = 0.00042), and the maximum AUC was 0.69 (Fig. 5b). Finally, we combined the training and validation set, calculated the risk score of the 502 patients according to the similar models, and performed survival analysis and plotted the survival curves of the high-risk group (n = 257) and low-risk group (n = 245) using the same threshold (Fig. 5c). As expected, LUAD patients in the high-risk group presented a worse prognosis (p < 0.0001), and the maximum AUC was 0.69.

To better apply the model to the actual clinical situation, clinical features of LUAD patients were introduced into the evaluations of the model and thus a composite nomogram to predict the survival probabilities of LUAD patients was constructed. By introducing the age, gender, tumor stage (I/II vs III/IV) and risk score as variables in the multivariate Cox regression analysis, and incorporating the above factors into the model as variables, a nomogram was obtained (Fig. 6a), in which, the line segment corresponding to each variable was marked with a scale that represented the value range of the variable, and the length of the line segment reflected the contribution of the factor to survival. The points in the nomogram represented the individual scores corresponding to each variable under different values and total points. The corresponding individual scores after introducing all the variables were summed to produce the total score. The last three lines represented the total score of 1-year, 3-year and 5-year survival rate. To verify the prognostic value of the nomogram model, the predictive ability of the nomogram was evaluated by tROC, and the AUC corresponding to 1-year, 3-year and 5-year survival was 0.698, 0.718 and 0.731, respectively (Fig. 6b). The prediction result of the C statistics on the nomogram model was 0.677 (95%CI, 0.633–0.721). The check chart indicated a good agreement between the predicted and actual results (Fig. 6c).

We downloaded the SNV data of 502 LUAD patients, and subgrouped based on the risk score. Each figure of the mutation landscape indicated the top 20 genes with the highest mutation frequency, and their mutations types were labeled with different colors (Fig. 7). The upper and right bar graph represented the total number of mutations in each sample, and the mutation frequency of each gene in all samples. It is revealed that genes with a higher mutation frequency in the high-risk group were consistent with those in the low-risk group, while their mutation frequencies were higher in the high-risk group [TP53 (53%), TTN (50%), MUC16 (42%), CSMD3 (40%), and RYR2 (39%)].

To further explore the differences in biological functions and pathways between the high-risk group and low-risk group, the DESeq2 package was used to analyze expression levels of DEGs between the high-risk group and low-risk group (logFC > 1, FDR < 0.05). DEGs screened out from both groups were later subjected to GO and KEGG enrichment analysis. The results of pathway enrichment and GO enrichment were shown in Fig. 8, and Additional file 7, 8: Table S3–S4. It is shown multiple pathways were enriched in the metabolism. In detail, the mainly enriched pathways included the neuroactive ligand-receptor interaction, metabolism of xenobiotics by cytochrome P450, steroid hormone biosynthesis, staphylococcus aureus infection pathway, IL-17 signaling pathway, retinol metabolic pathway, etc. The results of GO enrichment showed that the biological functions and processes of DEGs were mainly enriched in the keratinization, axoneme assembly, and microtubule bundle formation. Enrichment results could support the reasons for the variations in the prognosis of LUAD from a functional level. Moreover, some pathways and functions were identified closely correlated to ferroptosis or iron metabolism, which provided references for further research.

According to the threshold of risk score, the correlation between risk score and immune status in LUAD patients was further explored. We obtained a total of 29 immune-related gene sets involving multiple immune cells and immune-related functions or pathways. To quantify the enrichment degree of transcriptome data in the 29 immune gene sets, separate enrichment scores for each paring of a sample between the high-risk risk and low-risk group were estimated by the ssGSEA and compared by depicting a boxplot (Fig. 9). The enrichment scores of concentrations of immune cell genes in aDCs, iDCs, B-cells, mast cells, neutrophils, NK cells, T helper cells, Th2 cells and TIL were significantly different between the high-risk group and low-risk group (p < 0.05). Besides, the enrichment scores of these immune function gene sets in HLA, inflammation promoting, MHC class I, T cell co-stimulation and type II IFN response were also significantly different between groups (p < 0.05).