Identifying a cervical cancer survival signature based on mRNA expression and genome-wide copy number variations

Abstract

Cervical cancer mortality is the second highest in gynecological cancers. This study developed a new model based on copy number variation data and mRNA data for overall survival prediction of cervical cancer. Differentially expressed genes from The Cancer Genome Atlas dataset detected by univariate Cox regression analysis were further simplified to six by least absolute shrinkage and selection operator (Lasso) and stepwise Akaike information criterion (stepAIC). The study developed a six-gene signature, which was further verified in independent dataset. Association between immune infiltration and risk score was investigated by immune score. The relation between the signature and functional pathways was examined by gene set enrichment analysis. Ninety-nine differentially expressed genes were detected, and C11orf80, FOXP3, GSN, HCCS, PGAM5, and RIBC2 were identified as key genes to construct a six-gene signature. The prognostic signature showed a significant correlation with overall survival (hazard ratio, HR = 3.45, 95% confidence interval (CI) = 2.08–5.72, p < 0.00001). Immune score showed a negative correlation with the risk score calculated by the signature (p < 0.05). Four immune-related pathways were closely associated with risk score (p < 0.0001). The six-gene prognostic signature was an effective tool to predict overall survival of cervical cancer. In conclusion, the newly identified six genes may be considered as new drug targets for cervical cancer treatment.

Impact statement

We explored a new prognostic signature with six genes based on CNV data and mRNA data. The robustness of the six-gene signature in patient risk stratification, that is, categorization of patients into low- and high-risk groups, was confirmed. Moreover, risk score and immune score were closely associated. These six genes could be employed as a novel prognostic signature in predicting cervical cancer prognosis.

Introduction

Primary effective preventions of cervical cancer such as human papillomavirus (HPV) vaccination have been implemented long ago; cervical cancer incidence still ranks the fourth highest among women. ¹ In 2020, about 604,127 new cervical cancer cases were diagnosed all over the world. ¹ Recommended therapies have been proposed by the International Federation of Gynecology and Obstetrics (FIGO) based on the FIGO staging system. ² Although lymph node metastasis, tumor size, and invasion are key factors in deciding the recurrence risk of cervical cancer, cervical cancer patients evaluated as having a low risk of recurrence still develop cancer recurrence within two years of first diagnosis. ³ Therefore, an effective prediction system is needed to improve current personalized treatment.

A series of prognostic biomarkers have been discovered for survival prediction of cervical cancer patients, for instance, immune signatures,^4–6 a 10-gene prognostic model for aggressive cervical cancer, ⁷ biomarkers of long non-coding RNAs,^8–10 biomarkers of microRNAs,^11,12 biomarkers of histone genes, ¹³ and some other prognostic genes or pathways.^14–16 However, to date, a CNVs-based prognostic signature for cervical cancer has not been explored. According to previous studies, CNVs are closely associated with tumorigenesis in cancers such as lung cancer, leukemia, breast cancer.^17–20 According to the correlation between differential gene expression and CNVs, Shao et al. explored the relationship between CNVs and its downstream effect, particularly on tumor suppressor genes and oncogenes. ²¹ Advancement in DNA array technology has gradually uncovered small duplications or deletions related to cancers. Database of the Catalogue of Somatic Mutations in Cancer (COSMIC) collected 1,179,545 CNVs from previous literature (published August 2018) with tumor samples and genomic data. ²²

Considering close relationship between CNVs and cancer development, this study integrated CNV data and mRNA data to identify prognostic genes in cervical cancer. The objective of the present work was to develop a novel effective signature for overall survival prediction and assisting clinical treatments of cervical cancer. CNV and mRNA expression data were mined from available database through bioinformatics, and we demonstrated the relation between immune response and prognostic signature. In addition, the robustness of the prognostic signature developed in this study was compared with the existing ones.

Materials and methods

Data source and preprocessing

Cervical cancer samples with CNV data, mRNA data, clinical features, and patients’ survival data were collected from TCGA database (https://portal.gdc.cancer.gov/). GSE44001 profiles with mRNA data were downloaded from Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). We excluded samples without survival data or follow-up information. After preprocessing, 291 samples in TCGA dataset were included, and another 300 samples came from GSE44001 profiles (Supplementary Table S1). The workflow of constructing prognostic model of cervical cancer is shown in Figure 1.

Figure 1.

The flow diagram of developing a prognostic model of cervical cancer. (A color version of this figure is available in the online journal.)

TCGA: The Cancer Genome Atlas; CNV: copy number variation.

Identifying DEGS from TCGA dataset

CNVs were converted to genes by BEDTools, ²³ and genes with a CNV of |segment_mean| > 0.2 were retained for further analysis. After conducting Chi-square test, differential CNVs from tumor samples and normal samples were detected, and differential genes were outputted (p < 0.05). According to false discovery rate (FDR) < 0.01 and |log2(fold change, FC)| > 1, DEGs were screened from mRNA data using Limma R package.^24,25 Intersected DEGs between CNV and mRNA data were collected for further analysis.

Construction of prognostic model

TCGA dataset was randomly classified into training group and test group with a proportion of 1:1 using sample function in R with random seed 2119. The data showed no significant difference between two groups except for Grade item. GSE44001 served as a validation dataset. Univariate Cox regression analysis was conducted to analyze DEGs in the training dataset. For the purpose of model simplification, Lasso together with five-fold cross-validation and AIC were used to reduce variations and to construct the optimized model.^26,27 Risk score was defined as coefficient 1*mRNA expression of gene 1 + coefficient 2*mRNA expression of gene 2 +…coefficient n*mRNA expression of gene n. Then, the risk score was converted to z-score. The samples were classified into two risk groups (low-risk and high-risk) with z-score = 0 as the cut-off. The prognostic model was validated using test dataset and validation dataset. Group overall survival was assessed by plotting Kaplan-Meier survival curve. Survival analysis was conducted with log-rank test. p < 0.05 was defined as significant.

Calculation of immune score

Immune score was calculated by Estimation of STromal and Immune cells in MAlignant Tumors using Expression data (ESTIMATE) in the R package. ²⁸ ESTIMATE could deduce the fraction of stromal and immune cells in cancer samples based on gene expression signature. Additionally, Microenvironment Cell Populations-counter (MCP-counter) was employed to quantify abundance score of various immune-related cells, including B lymphocytes, CD3⁺ T cell, neutrophils, CD8⁺ T cell, endothelial cell, NK cells, fibroblast, monocytic lineage, cytotoxic lymphocyte, and myeloid dentritic cell. ²⁹

Gene enrichment analysis

WebGestalt R package was employed to analyze GO (gene ontology) function and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway of DEGs.^30,31 Relation between KEGG pathways and risk score was analyzed by GSVA R package.^32,33 Functional pathways showing a correlation coefficient > 0.3 were seen as correlated with the risk score. Statistical differences in different pathways between two risk groups were analyzed by Wilcoxon test. Significant results were defined with a p < 0.05.

Results

Screening DEGs based on CNV and mRNA datasets

In CNV dataset, we screened 8413 genes showing differential CNVs through comparing the CNV of cervical cancer samples with the normal samples using Chi-square test (p < 0.05). In mRNA dataset, Limma in the R package was used to detect DEGs. A total of 2000 DEGs incorporating 961 down-regulated and 1039 up-regulated genes were filtered with the conditions of FDR < 0.01 and |log₂FC| > 1 (Figure 2(a)). Seven hundred and seventy-four DEGs incorporating 359 down-regulated genes and 405 up-regulated genes were identified from genes showing CNVs and mRNA dataset (Figure 2(b)). Furthermore, the enrichment analysis of 764 genes in GO function showed that these genes were related to seven molecular function terms, 148 biological process terms, 48 cellular component terms (FDR < 0.05). Among these terms, “condensed chromosome kinetochore,” “DNA-dependent DNA replication,” and “DNA replication origin binding” were the most enriched terms of cellular components, biological processes, and molecular functions, respectively (Figure 3(a) to (c)). In KEGG pathways, seven pathways, including “p53 signaling pathway,” “Oocyte meiosis,” “DNA replication,” “Fanconi anemia pathway,” “Cell cycle,” “One carbon pool by folate,” and “Homologous recombination” were found to be enriched to these genes (FDR < 0.05, Figure 3(d)).

Figure 2.

Screened DEGs from mRNA data of TCGA dataset. (a) Volcano plot of DEGs (1039 up-regulated genes and 961 down-regulated genes). Tumor/normal represents the expression ratio of tumor samples and normal samples. Down means down-regulated samples (blue) and up means up-regulated samples (red). None means no significant change of expression (grey). (b) Venn plot within down-regulated genes (violet), up-regulated genes (red), and differential CNVs converted as genes (green). (A color version of this figure is available in the online journal.)

FDR: false discovery rate; FC: fold change; CNV: copy number variants.

Figure 3.

Gene expression enrichment of DEGs in GO function and KEGG pathways. (a) The top 10 biological process terms. (b) The top 10 cellular component terms. (c) Seven molecular function terms. (d) Seven KEGG pathways. Size means counts of genes in different terms. (A color version of this figure is available in the online journal.)

FDR: false discovery rate.

Constructing a prognostic model of cervical cancer

The samples of TCGA dataset were randomly grouped into training dataset (146 samples) and test dataset (145 samples). The results did not show any significant difference between the two groups except for Grade item (G1 to G4, and GX, p > 0.05, Supplementary Table S2). The survival coxph function in the R package and univariate Cox regression analysis showed that 764 DEGs and outputted 99 genes were overall survival related (p < 0.05). For a further simplification of the gene set with high accuracy, Lasso analysis was used here to reduce gene numbers. Genes showing a coefficient close to zero with the increasing lambda (λ) value were retained to construct a model with the least number of genes or variables (Figure 4(a)). The confidence interval of λ is shown in Figure 4(b). When λ = 0.059, the model had 12 genes. Then, we used stepAIC R package to further optimize the model. Finally, a total of six genes, namely, C11orf80, FOXP3, GSN, HCCS, PGAM5, and RIBC2, were enrolled in our prognostic model. The formula of risk score was as follows

$$\text{Risk Score}=\text{1}.\text{256}\times C11orf80-0.\text{846}\times \mathit{FOXP}3-0.\text{489}\times \mathit{GSN}-0.\text{553}\times \mathit{HCCS}-\text{1}.0\text{81}*\mathit{PGAM}5-0.\text{473}\times \mathit{RIBC}2$$

Figure 4.

Optimization of prognostic model by Lasso. (a) Coefficient variation following with λ variation of 99 differentially expressed genes in Lasso model. One curve represents one gene. Dashed line means the site when λ = 0.059. (b) Partial likelihood deviance of λ (green and blue line). Confidential interval was shown in black line. Red dot means the site when λ = 0.059. (A color version of this figure is available in the online journal.)

The six genes were distributed in chromosome 9 (GSN), chromosome 11 (C11orf80), chromosome 12 (PGAM5), chromosome 22 (RIBC2), and chromosome X (HCCS and FOXP3) (Supplementary Figure S1). Compared with normal samples, cervical cancer samples had higher expression of the six genes (p < 0.05, Supplementary Figure S2). The curve of Kaplan-Meier survival demonstrated that the training set samples were divided into high- and low-risk groups by FOXP3, GSN, HCCS and RIBC2 (p < 0.05), but the samples were not sensitive to the other two genes C11orf80 and PGAM5 (Supplementary Figure S3). Sixty-eight out of 146 samples in the training set were grouped into high-risk and the remaining 78 samples were grouped into low-risk group through calculating sample risk score (Figure 5). The mRNA expression of C11orf80 was higher in high-risk group, whereas the remaining five genes showed a higher expression in low-risk group (z-score of mRNA expression = 0 was selected as the cut-off for defining relatively low and high gene expression, Figure 5(a)). From Kaplan-Meier survival curve, it was shown that low-risk patients tended to live longer than high-risk patients (HR = 5.3, 95%CI = 3.21–8.75, p < 0.0001). Moreover, the robust effectiveness of the prognosis model in overall survival prediction of cervical cancer patients was confirmed by the area under curve (AUC), as AUC for five years, three years, and one year was 0.83, 0.82, and 0.89, respectively (Figure 5(b) and (c)).

Figure 5.

Performance of six-gene prognostic signature in the training dataset. (a) Classification of low-risk group (blue) and high-risk group (red) by risk score which was converted as z-score. The distribution of dead and alive samples in two risk types was shown under risk types. Association between risk types and expression of six genes. (b) ROC curves of one-year, three-year, and five-year survival with an AUC of 0.89, 0.82, and 0.83, respectively. (c) Kaplan-Meier survival curve of low-risk and high-risk groups (p < 0.0001, HR = 5.3, 95%CI = 3.21–8.75). (A color version of this figure is available in the online journal.)

CI: confidential interval; HR: hazard ratio.

Validating the six-gene prognostic signature

Prediction ability of the six-gene prognostic signature was evaluated using the test dataset from TCGA and validation dataset from GSE44001. In the test dataset, 145 samples were grouped into low-risk group (73 samples) and high-risk group (72 samples) (p < 0.001, Supplementary Figure S4). The accuracy of the signature in evaluating five-year, three-year, one-year overall survival was shown by an AUC of 0.84, 0.71, and 0.78, respectively (Supplementary Figure S4(b)). Similarly, the signature was robust in classifying a total of 291 samples from TCGA into two risk groups (p < 0.0001, Supplementary Figure S5). In the validation dataset, 144 out of 300 samples were predicted as having a low risk, while the rest 156 samples were evaluated as having a high risk (p < 0.05, Figure 6). One-year, three-year, and five-year AUC of overall survival was 0.73, 0.64, and 0.63, respectively (Figure 6(b)).

Figure 6.

Performance of the six-gene prognostic signature in the validation dataset (GSE44001). (a) Classification of low-risk group (blue) and high-risk group (red) by risk score which was converted as z-score. The distribution of dead and alive samples in two risk types was shown under risk types. Association between risk types and expression of six genes. (b) ROC curves of one-year, three-year, and five-year survival with an AUC of 0.73, 0.64, and 0.63, respectively. (c) Kaplan-Meier survival curve of low-risk and high-risk groups (p = 0.04, HR = 1.55, 95%CI = 1.09–2.19). (A color version of this figure is available in the online journal.)

CI: confidential interval; HR: hazard ratio.

Correlation analysis between risk score and clinical features

Based on TGCA dataset, we also analyzed correlation between risk score and clinical features. Distribution analysis of different clinical features in the two risk groups showed that the mortality was noticeably higher in high-risk group (p < 0.05, Figure 7(a)). Significantly different distribution of N stage, stage I to IV, and grade was detected between two risk groups (p < 0.05, Figure 7(c), (e), (g)), while there was no difference in T stage, M stage or age (Figure 7(b), (d), (f)). Survival analysis also showed a higher overall survival of low-risk group with different clinical features (Figure 8). Although no significant difference in N0 stage or M1 stage was detected, the risk score was closely associated with patients’ overall survival in most clinical features, including stage I to IV, M0 stage, N1 stage, T stage, and grade 1 to 4, age (p < 0.05, Figure 8). Univariate Cox regression analysis revealed that risk was closely related to overall survival (95%CI = 2.08–5.72, HR = 3.45, p < 0.00001), and that M stage, N stage, T stage were positively related to overall survival (HR > 1, p < 0.05) (Figure 9(a)). From the data of multivariable Cox regression analysis, overall survival was significantly related to risk type (HR = 2.52, 95%CI = 0.89–7.11, p < 0.00001) (Figure 9(b)).

Figure 7.

Comparison of different clinical features including survival status (a), T stage (b), N stage (c), M stage (d), stage I to IV (e), age (f), and grade 1 to 3 (g) in low-risk and high-risk groups. Group differences were analyzed by ANOVA. *p < 0.05. (A color version of this figure is available in the online journal.)

Figure 8.

Kaplan-Meier survival curve of different clinical features including age (a–b), T stage (c–d), N stage (e–f), M stage (g–h), stage I to IV (i–j), grade 1 to 4 (k–l) in low-risk and high-risk groups. Log-rank test was performed. p < 0.05 was considered to be significant. (A color version of this figure is available in the online journal.)

Figure 9.

Relation between risk score and clinical features. (a) Univariate Cox regression analysis of clinical features and risk score. (b) Multivariate Cox regression analysis of risk score and clinical features. Green diamond means the media of HR. (A color version of this figure is available in the online journal.)

CI: confidential interval; HR: hazard ratio.

A nomogram for predicting overall survival

We combined T stage and risk score to develop a nomogram based on TCGA dataset. From Figure 10(a), total points of risk score and T stage could be calculated, and mortality in five years, three years, and one year is shown. The predicted overall survival was corrected by the actual overall survival (Figure 10(b)). We conducted decision curve analysis (DCA) to evaluate overall survival prediction performance of the nomogram, T stage and risk score, and observed a high prediction accuracy of the nomogram (Figure 10(c)).

Figure 10.

Application of risk score in predicting overall survival. (a) A nomogram to predict one-year, three-year, and five-year death rate. Risk score and T stage were included to calculate total points. (b) The relation between predicted OS by nomogram and observed OS of one year, three years, and five years. (c) DCA plot of nomogram (red), T stage (green), and risk score (blue). (A color version of this figure is available in the online journal.)

OS: overall survival.

Correlation analysis between immune score and risk score

Immune score could act as a prognostic factor; therefore, its relationship with between risk score was assessed. Immune score of all the TCGA samples was calculated with the ESTIMATE in the R package. Figure 11 shows the difference of immune score in low- and high-risk groups. Low-risk patients showed higher Immune score, Stromal score, and ESTIMATE score than high-risk ones (p < 0.001, Figure 11). Furthermore, we evaluated immune infiltration of tumor through MCP-counter. Consistent with the ESTIMATE results, the immune score calculated by MCP-counter showed a higher enrichment of most immune cells (CD3⁺ T cells, endothelial cells, B lymphocytes, CD8⁺ T cells, NK cells, myeloid dentritic cells, cytotoxic lymphocytes, fibroblasts, and neutrophils, monocytic lineage) in low-risk group (p < 0.05, Figure 12). These two methods, for analyzing the distribution of immune cells, revealed a lower immune infiltration in high-risk group than low-risk group, suggesting that a worse survival was associated with a lower level of immune infiltration.

Figure 11.

Comparison of immune score between low-risk group (red) and high-risk group (blue) in the TCGA dataset. Student’s t test was employed to investigate the difference between two groups. ***p < 0.001. (A color version of this figure is available in the online journal.)

Figure 12.

Analyzing the immune score of different immune-related cells by MCP-counter in CD8⁺ T cells, cytotoxic lymphocytes, CD3⁺ T cells (T cells), NK cells, myeloid dentritic cells, monocytic lineage, B lymphocytes (B lineage), endothelial cells, fibroblasts, and neutrophils. The difference between two groups was examined by Student’s t test. –p > 0.05, *p < 0.05, **p < 0.01, ***p < 0.001. (A color version of this figure is available in the online journal.)

Identifying functional pathways related to risk score

To explore the relation between functional pathways and risk score, we conducted single sample gene set enrichment analysis (ssGSEA) analysis to calculate ssGSEA score of TCGA dataset samples. An Association study between risk score and ssGSEA score detected 16 pathways in negative correlation with risk score (Figure 13(a)); here, four immune-related pathways, namely, natural killer cell mediated pathway, B cell receptor signaling pathway, T cell receptor signaling pathway, and chemokine signaling pathway, were negatively related to risk score (p < 0.0001, Figure 13(b) to (e)).

Figure 13.

Correlation between risk score and KEGG pathways evaluated by ssGSEA. (a) Sixteen functional pathways negatively related to risk score. (b) Comparison of high- and low-risk groups in four immune-related pathways (natural killer cell-mediated pathway, chemokine signaling pathway, and B cell receptor signaling pathway, and T cell receptor signaling pathway). Wilcoxon test was performed. (A color version of this figure is available in the online journal.)

Comparison with other prognostic signatures

We also compared the current gene signature with previously developed ones. We selected the gene signature for cervical cancer developed in recent two years. Considering that the number of prognostic genes can significantly affect AUC, we included four signatures with similar gene numbers compared with our signature. Therefore, we selected four prognostic signatures of cervical cancer, specifically, a five-gene signature (GALNTL6, ARSE, DPAGT1, GANAB, and FURIN) of Ju et al., ³⁴ a six-gene signature (ANGPTL4, GOT1, HSPA5, IER3, PFKM, and PFKFB4) by Cai et al., ³⁵ a five-gene signature (EPHX2, CHAF1B, KIAA1524, CDC45, and RMI2) by Liu et al., ³⁶ an eight-gene signature (CCDC136, ABCG2, CYP26A1, TNNI3, SYT13, FOXC2, CXCL5, and TMEM233) by Xie et al. ³⁷ To ensure the comparability, the same methodology was applied to determine the risk score of 291 samples in TCGA dataset using the four signatures. The results showed that all the samples were clearly categorized into low- and high-risk groups by the four signatures (p < 0.05, Figure 14). The receiver operating characteristic (ROC) demonstrated that the five-gene signature developed by Liu et al. had the highest AUC of 0.76 in predicting one-year survival (95%CI = 0.65–0.88, Figure 14(e)), and that the five-gene signature by Cai et al. had the highest AUC of 0.72 in predicting three-year survival (95%CI = 0.64–0.81, Figure 14(c)); moreover, the five-gene signature of Ju et al. showed the highest AUC of 0.75 in predicting five-year survival status (95%CI = 0.65–0.84, Figure 14(a)). As the five-gene signature of Ju et al. had a higher AUC of five-year survival than our signature, we further calculated C-index of these two signatures. Our signature was found to have a relatively higher C-index value in all three datasets (training dataset, test dataset, and validation dataset) than Ju et al.’s signature, suggesting that our six-gene signature was more effective in predicting overall survival of cervical cancer patients (Supplementary Figure S6). Compared with an AUC of 0.71, 0.72, 0.79 in predicting five-year, three-year, and one-year survival of our six-gene signature, respectively, the overall effectiveness of these four signatures for predicting survival was all less preponderant than the prognostic signature in this study (Supplementary Figure S5).

Figure 14.

The performance of other prognostic signatures of cervical cancer from literature. (a–b) ROC and Kaplan-Meier survival curve of five-gene signature from Ju et al. HR = 1.76, 95%CI = 1.37–2.24, p = 0.022. (c–d) ROC and Kaplan-Meier survival curve of six-gene signature from Cai et al. HR = 2.15, 95%CI = 1.62–2.85, p < 0.0001. (e–f) ROC and Kaplan-Meier survival curve of five-gene signature from Liu et al. HR = 2.06, 95%CI = 1.61–2.64, p = 0.00037. (g–h) ROC and Kaplan-Meier survival curve of eight-gene signature from Xie et al. HR = 1.96, 95%CI = 1.57–2.46, p < 0.0001. (A color version of this figure is available in the online journal.)

HR: hazard ratio; CI: confidence interval.

Discussion

Studies have previously developed prognostic signatures of cervical cancer based on long non-coding RNAs, microRNAs, mRNAs, immune genes, histone genes; however, signatures developed based on CNVs have not been explored. The study established a prognostic model based on CNV data along with mRNA data of cervical cancer samples from GEO and TCGA databases for the characterization and prediction of overall survival of cervical cancer. This was the first study that combined CNV data and gene expression data to identify differential genes. We detected 99 DEGs from TCGA dataset and further simplified the gene set using Lasso and AIC analysis. Most genes in the six-gene signature (C11orf80, FOXP3, GSN, HCCS, PGAM5, and RIBC2) could characterize patients’ overall survival, and the prediction system showed a robust performance in the independent dataset (GSE44001). Further analysis between overall survival and the prognostic signature revealed that the risk score was relevant to immune response, with low-risk group showing a high immune score (Figures 11 and 12), indicating that one or several prognostic genes were involved in immune-related process.

Of these six prognostic genes, FOXP3, GSN, and PGAM5, were closely involved in the relation between immune score and risk score. As a transcriptional regulator, FOXP3 functions essentially in activating or suppressing the function of regulatory T-cells through interacting with cytokines and other some transcription factors, NF-kappaB and nuclear factor of activated T cells (NFAT), for instance. ³⁸ It has been reported that FOXP3 is strongly related to many cancers, including colorectal cancer, non-small lung cancer, papillary and follicular thyroid cancers, breast cancer, prostate cancer, melanoma, gastric cancer, pancreatic cancer, and cervical cancer.^39,40 In this study, consistent with previous researches, higher expression of FOXP3 has been detected in tumor samples in comparison with normal ones.^41,42

Gelsolin (GSN), which participates in immune response, is also involved in macrophage recruitment, phagocytosis, and could exogenously enhance the function of T-lymphocytes.^43,44 Lower expression of GSN in tumor samples was detected; interestingly, GSN also had lower expression in high-risk group, indicating a lower immune response in high-risk group and higher tendency of developing tumor progression. Furthermore, it was also found that GSN has tumor-suppressive effect on bladder cancer and breast cancer.^45,46 Consistently, Abedini et al. discovered that overexpressed GSN was significantly related to aggressive gynecological cancers. ⁴⁷ It has been demonstrated that NKT cell activation was mediated by the mitochondrial phosphatase phosphoglycerate mutase 5 (PGAM5), and RIPK3-PGAM5-DRP1 signaling axis could mediate the crosslink between host immunity and mitochondrial function in inflammatory diseases. ⁴⁸

Interestingly, RIBC2 has been previously determined as a prognostic gene of cervical cancer through bioinformatics analysis.^49,50 Higher expressed RIBC2 were found to be positively correlated with a more favorable overall survival, which was consistent with our results (Supplementary Figure S3, Figures 8 and 9). However, its function and biological process in tumorigenesis was still unclear. It was indicated that RIBC2 interacts with YBX1, which is associated with cancer proliferation in numerous tissues ⁵¹ and also with TREX1, aberrant mutations of which are involved in immune-related diseases. ⁵² The involvement of C11orf80 and HCCS in cervical cancer or in other cancers has not been reported; thus, their mechanism and function in oncogenesis should be further investigated.

Compared with other prognostic signatures, our six-gene signature based on CNV and mRNA data showed higher AUC and C-index, indicating that it was more effective in predicting overall survival. Although the methodology used in the present study seemed to be similar to Ju et al., we focused more on CNVs, while Ju et al.’s study was based on post-translational modifications (PTMs). CNVs can affect gene expression. In clinical practice, CNVs are more stable and easy to be detected than the variation of gene expression. Detecting the expression of differential genes formed a part of our analysis, but it should be clear that the six-gene signature in our study (C11orf80, FOXP3, GSN, HCCS, PGAM5, and RIBC2) was not overlapped with the five-gene signature developed by Ju et al. (GALNTL6, ARSE, DPAGT1, GANAB, and FURIN), as two studies used different strategies when identifying prognostic genes. In addition, our study used three datasets (training dataset, test dataset, and validation dataset) to construct and validate the prognostic model, showing a higher reliability than Ju et al.’s study in which only training dataset and test dataset were used. Although screening DEGs from both CNVs and mRNA data can eliminate the effect of false DEGs, using this methodology may omit some certain DEGs at the same time. In addition, as our results were obtained only based on bioinformatics analysis, future validation with more experimental data and clinical patients is also required.

In conclusion, the study developed a novel prognostic signature with six genes according to CNV data and mRNA data. The robustness of the six-gene signature to stratify patients into high- and low-risk groups was confirmed. Moreover, the current findings revealed a close correlation between risk score and immune score, providing new insights into development of new drug targets for treating cervical cancer. In clinical practice, these six genes could be a novel prognostic signature to predict prognosis of cervical cancer patients.