Construction and validation of a prognostic model of RNA binding proteins in clear cell renal carcinoma

Human Embryonic Renal Cell Line (293 T), Renal non-cancer cell line (HK2), Human renal carcinoma cell lines 786-O and OS-RC-2 were obtained from the Typical Culture Preservation Commission Cell Bank, Chinese Academy of Sciences (Shanghai, China). Materials for the cell culture process, including Fetal Bovine Serum (FBS), RPMI 1640 culture medium, trypsin, penicillin and streptomycin, were purchased from Gibco (Grand Island, NY, USA). The culture medium for all cell lines contained 90% RPMI 1640, 10% FBS and 1% antibiotics (100 U/ml streptomycins and 100 U/ml penicillin). All cell lines were cultured at 37 °C, 5% CO₂.

Each cell was inoculated at a density of 1 × 10⁶ cells/well in 6-well plates, 2 ml of mixed media per well and incubated for 24 h at 37 °C in a humidified incubator with 5% CO₂. Total RNA was extracted from the 6-well plates using Trizol reagent (Vazyme, China).

Twenty pairs of clear cell renal cancer tissues and paraneoplastic tissue specimens were obtained from patients after radical surgery for clear cell renal cancer at the Affiliated Hospital of Medical College of Qingdao University. The detailed clinicopathological characteristics of the 20 ccRCC patients are shown in Table 1. The cancerous and paraneoplastic tissues (5 cm apart) were rinsed in sterile PBS and rapidly frozen in liquid nitrogen within 30 min after removal. The Institutional Review Board approved the study protocol, and informed consent was obtained from the patients.

We obtained all transcriptome (FPKM) and clinical data of ccRCC from the TCGA (https://​portal.​gdc.​cancer.​gov/​) database, it contains 72 paraneoplastic samples and 539 tumor samples. In human cells, 1542 RBPs genes (Supplement Table 1) were screened by high-throughput screening [7]. We extracted 1495 RBPs genes expression data from TCGA data. The process of this study follows the flow chart (Fig. 1).

To observe whether RBPs genes are expressed differently in tumor and normal tissues. In R (Version 3.6.2) language environment, “limma" packages (http://​www.​bioconductor.​org/​packages/​release/​bioc/​html/​limma.​html) were used for data correction, and Wilcox test was used for data difference analysis. The cut-off value of our screening criteria was |logFC|> 1 and FDR < 0.05 (Adjusted p-value). Heat maps were drawn using “pheatmap” package, and volcanoes were mapped. Finally, we collate the differential expression RBPs (DERBPs) output into a readable file.

To further understand the biological function of DERBPs in ccRCC. We carried out Gene Ontology (GO) [17] enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) [18] pathway enrichment analysis of DERBPs. A p-value < 0.05 and FDR < 0.05 was used as the GO and KEGG enrichment analysis filtration standard. The analysis process is carried out in the R language environment. The R packages used include "clusterProfiler", "ggplot2" and" enrichplot".

To explore the interaction between DERBPs. We used STRING (http://​www.​string-db.​org/​, Version 11.0), an online analysis tool, for PPI analysis [19]. Visualise the PPI network using Cytoscape (Version 3.7.2).

To further verify the model's prediction efficiency, samples were randomly divided into training group and test group in a 1:1 ratio. We obtained information on a total of 537 patients from the TCGA database. A total of 530 patient sample data were obtained by excluding samples with missing data. The training group contained 266 samples, while the test group contained 264 samples. We screened the prognosis genes associated with ccRCC survival from the PPI network. We performed univariate Cox regression analysis of genes in the network to identify genes associated with overall survival (p < 0.0001). Finally, we performed multivariate Cox regression analysis of genes related to survival, constructed a risk prediction model of RBPs, and calculated the patients' risk scores in the training group. Risk score = Ʃ (β_n × Exp_n). In the formula, β represents the regression coefficient, and Exp represents the expression level of related genes. All operations for this step are performed in the R language environment. The R package used in the operation procedure includes "survival", "caret", "glmnet" and "survminer".

We divided the 264 samples in the test group into high-risk and low-risk groups based on the median risk score of the model. The overall survival of the two groups was observed. We also draw the receiver operating characteristic (ROC) curve to evaluate our prediction effectiveness and by using the "survivalROC” package in R.

To further investigate the relevance of the model to clinical practice. In the R environment, the chi-square test is used for analysis. We were able to assess the correlation between the model and clinicopathological features. We also performed an independent predictive analysis of the model to determine whether our model can be used as a prognostic indicator of ccRCC alone. To better apply the clinical practice model, we drew a nomogram to predict patients' survival status accurately.

To further analyze the accuracy of the model, we validated the expression levels of model-related genes at the tissue and cell line levels, respectively. The procedure for tissue extraction of RNA is provided in Supplementary Material (S1). For cells, total RNA was extracted from the 6-well plates using Trizol reagent (Vazyme, China). A microgram of total RNA was reverse transcribed to cDNA with HiScript III RT SuperMix for qPCR Reverse Transcriptase (Vazyme, China). Quantitative real-time PCR was performed to analyse the cDNA according to the SYBR Color qPCR Master Mix (Vazyme, China) instructions with a Roche LightCycler 480II real-time PCR detection system (Roche, Switzerland). Normalization was carried out using GAPDH. Details of the primers used in this experiment are given in Table 2.

The data mining and processing in this study were done using R software (version 3.6.2). All the experimental data (PCR data) in this study have been repeated 3 times and are expressed as \(\overline{\mathrm X}\) ± S. The experimental data (PCR data) were all statistically analyzed using SPSS software (version 26.0) and the data were visualised using Graphpad Prism software. For the subgroups in this experiment where only two groups were compared, we used the t-test for comparison, and for data containing more than two subgroups, we used one-way ANOVA for statistical analysis. The analysis results p < 0.05 considered the differences to be statistically significant.

In R language environment, 1495 RBPs-related genes in ccRCC were analyzed according to the screening criteria (|log FC|> 1, FDR < 0.05). The RNA expression matrix of 72 paracancer samples and 539 tumor samples were analysed for differential expression, statistically using the wilcoxon test. The results showed 125 dysregulated RBPs in tumor tissues, among which 38 genes were up-regulated, and 87 were down-regulated (Supplement Table 2). We visualized the differentially expressed genes by mapping heat map and volcanic maps (Fig. 2).

We carried out GO and KEGG enrichment analysis for DERBPs. GO enrichment analysis includes three categories, including biological process (BP) analysis, cell component (CC) analysis and molecular function (MF) analysis. GO enrichment results showed that the enrichment of most genes was related to RNA metabolism and protein formation. Such as: “regulation of mRNA metabolic process," "regulation of RNA splicing," "RNA catabolic process," "RNA phosphodiester bond hydrolysis” and "RNA splicing". Figure 3A shows the GO enrichment results, with each function showing only the top 10 critical terms. KEGG enrichment analysis results showed that it was related to "Ribosome," "RNA transport," and "mRNA surveillance pathway." The analysis results are shown in Fig. 3B.

To explore the interaction between DERBPs, we used STRING, an online web page analysis tool, to build a PPI network (Confidence = 0.7). Which is visualized by using Cytoscape (Fig. 3C). Our PPI network contains 225 networks and 100 nodes. We only show the first 30 nodes based on the number of genes connected (Fig. 3D, Supplement Table 3).

We performed univariate Cox regression analysis of RBPs in the network to screen genes associated with survival. The results showed that 17 genes (p < 0.0001) were correlated with the survival of ccRCC (Fig. 3E, Supplementary Fig. 1).

In the training group, we used the built-in functions in R to perform a multivariate Cox regression analysis of 17 genes associated with survival, assessing the relative effect of each gene, comparing the regression p-values of individual genes, resulting in the identification of a four RBPs-related (U2AF1L4, RPL22L1, EZH2, RNASE2) prediction model (Fig. 4, Table 3). The patient's risk score was calculated according to the model, and the risk formula was as follows: Risk score = (EXP_U2AF1L4 Χ 0.4984) + (EXP_RPL22L1 X 0.5104) + (EXP_EZH2 X 0.4014) + EXP_RNASE2 X 0.6551). We divided patients into high-risk and low-risk groups based on the median risk score in the training data set. Kaplan–Meier survival analysis was performed to see the significance of survival between the high and low risk groups. We plotted survival curves to assess survival differences between the two groups, and we plotted the ROC curve to see the accuracy of our model predictions. The results showed that the five-year survival rate for the high-risk group was 40.2% (95% CI = 0.313 ~ 0.518), while the five-year survival rate for the low-risk group was 84.3% (95% CI = 0.767 ~ 0.926). The five-year survival rate for the low-risk group was much higher than that of the high-risk group (Fig. 5A). The ROC curves (AUC = 0.748) also showed that our model had stable predictive power (Fig. 5B).

To verify the predictive power of the model, we validated our model in the test group. Like the above risk formula, patients in the verification group were divided into high-risk and low-risk groups according to the median risk score. The survival and prognosis of the two groups were observed. The results in Fig. 5C showed that the high-risk group had a worse prognosis than the low-risk group. The five-year survival rate for the high-risk group was 47.9% (95% CI = 0.381 ~ 0.603), while the five-year survival rate for the low-risk group was 73% (95% CI = 0.639 ~ 0.834). The ROC analysis showed an AUC = 0.690, also demonstrating that our model also had good predictive power in the test group (Fig. 5D).

This analysis was all compared using t-tests. The analysis results p < 0.05 considered the differences to be statistically significant. The RT-qPCR showed that the mRNA expressions of EZH2, RPL22L1, RNASE2, and U2AF1L4 were higher in renal tumor tissues than in normal renal tissues (Fig. 6A). In addition, this pattern was verified at the cellular level that the expressions of EZH2, RPL22L1, RNASE2, and U2AF1L4 were higher in renal tumor cells (786-O, OSRC) than in non-cancerous renal cells (HK2) (Fig. 6B). Collectively, these results demonstrated the accuracy of our model.

To further evaluate whether our model can be used as a sole factor in assessing prognosis of ccRCC. Univariate and multivariate Cox regression analyses were performed for the model and clinical indicators. Univariate analysis (Fig. 7A) showed that patient age, tumor stage, tumor grade, and our model were associated with the prognosis of ccRCC patients (p < 0.001). Multivariate analysis (Fig. 7B) showed that patient age, tumor stage, and our model were associated with prognosis in ccRCC patients (p < 0.001). The results altogether demonstrated that our risk model could be used as an independent prognostic factor for ccRCC, and the accuracy of prediction is not affected by any other clinicopathological indicators. Chi-square test was used to analyze the correlation between the model and clinicopathological features. The results showed that the high-risk score was closely related to tumor stage (p < 0.001), pathological grade (p < 0.001), metastasis status (p < 0.01), but not related to gender or age (Fig. 7C). Also, the relationship between the expression levels of four genes in the model and the risk score was demonstrated. To better apply the model in the clinic, we developed the nomogram of the model. The 1-year, 3-year, and 5-year survival rates for ccRCC patients were predicted using the nomogram (Fig. 8).