Construction and validation of a novel prognostic risk model in cervical cancer: integrating lactylation-related genes linked to radiotherapy‑associated transcriptional changes

Abstract

Background

The association between lactylation and tumor radiotherapy has become a widely studied hotspot. However, the prognostic value of lactylation-related genes (LRGs) linked to radiotherapy‑associated transcriptional changes in cervical cancer remains unclear.

Methods

Based on gene expression data from public databases and LRGs, we identified key genes for cervical cancer prognostics and verified their expression levels. Subsequently, LASSO regression with 10-fold cross-validation was used for model construction and internal validation. Patients were divided into different groups according to the optimal cut-off value of risk score, and their differences in biological function/pathway, immunoregulation, and tumor mutation burden (TMB) were analyzed. A nomogram for predicting the survival probability was also constructed.

Results

Our study identified two key genes (RFC4 and STMN1), whose expression was upregulated in tumor tissues but significantly downregulated during radiotherapy. The risk model incorporating the two genes was verified internally and exhibited favourable prognostic stratification capability. Patients were divided into high- and low-risk groups; differentially expressed genes between the groups were significantly enriched for immune-, angiogenesis-, and metabolism-related functions/pathways. Furthermore, the two groups exhibited distinct immune cell infiltration levels and gene mutation, with the low-risk group showing superior survival outcomes. Multivariate Cox analysis identified risk score, overall stage, and T stage as independent prognostic factors (P < 0.05). The nomogram incorporating these factors effectively predicted 1-year, 2-year, and 3-year survival rates.

Conclusion

This study provides a promising prognostic model for risk stratification of cervical cancer patients and identifies two biomarkers, which will help promote the understanding of the relationship between radiotherapy, lactylation, and prognosis in cervical cancer.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05351-z.

Introduction

Cervical cancer is one of the most common malignant tumors among women worldwide, and its morbidity and mortality are among the highest in gynecological tumors [1]. Despite the remarkable progress in screening and diagnostic techniques for cervical cancer in recent years, a considerable number of patients tend to be at the middle or advanced stage when diagnosed [2, 3]. Radiotherapy is the primary treatment for locally advanced cervical cancer, but some patients may develop radiotherapy resistance or recurrence, leading to treatment failure [4, 5]. Currently, clinicians mainly rely on the clinicopathological characteristics of cervical cancer patients, such as tumor stage, pathological type, and lymph node metastasis, to assess patients’ prognosis and formulate treatment plans [6–8]. However, these traditional indicators are insufficient in classifying patients’ risk stratification. Therefore, it is of great significance to search for reliable prognostic markers for cervical cancer risk prediction and radiotherapy guidance.

Post-translational modifications (PTMs) are pivotal in tumorigenesis by regulating protein function and cellular signaling, but remain understudied due to their high dynamics and technical limitations in identification [9, 10]. Meanwhile, they represent promising targets for cancer therapy given their extensive roles in malignant behaviors and treatment resistance [9]. To date, PTMs such as phosphorylation, acetylation, and ubiquitination in cervical cancer have been extensively studied [11]. However, the active glycolytic metabolism in the tumor microenvironment leads to a massive lactate accumulation, making lactylation, which uses lactate as the substrate, a crucial bridge linking metabolic reprogramming and malignant phenotype [12]. This emerging PTM promotes tumor progression and affects treatment outcome through metabolic reprogramming, epigenetic regulation, and immunomodulation [13, 14]. Accumulating evidence underscores its significant role in radiotherapy response. For instance, MRE11 K673 lactylation mediated radiation resistance in triple-negative breast cancer [15]. Histone lactylation transcriptively activated MCM7 to enhance the radio-resistance of hepatocellular carcinoma [16]. In cervical cancer, Meng et al. revealed that the DCBLD1 lactylation inhibited self-ubiquitination degradation and activated the pentose phosphate pathway, thereby promoting cervical cancer cell proliferation [17]. Huang et al. found that the H3K18 lactylation upregulated GPD2 transcription, driving M2 macrophage polarization and promoting malignant progression of cervical cancer [18]. Notably, these studies neither further investigated the association between lactylation and radiotherapy in cervical cancer nor established a gene signature capable of predicting patient survival outcomes. Therefore, developing a risk model specifically targeting lactylation-related genes (LRGs) linked to cervical cancer radiotherapy would provide valuable insights for advancing treatment strategies for this disease.

In this study, we integrated cervical cancer gene expression profiles with LRGs and identified key genes in the radiotherapy cohort datasets. Using these genes, we constructed and internally validated a prognostic risk model and investigated its correlation with tumor characteristics, immune regulation, and survival prognosis. This model shows potential as a prognostic tool for risk stratification and survival evaluation in cervical cancer.

Materials and methods

Data source and processing

The RNA-seq data (TCGA-CESC.star_fpkm.tsv) of The Cancer Genome Atlas-cervical cancer (TCGA-CESC) and its matched survival information (TCGA-CESC.survival.tsv) and clinicopathological characteristic data (TCGA-CESC.clinical.tsv) were obtained from the Xena database (https://xena.ucsc.edu/) of the University of California, Santa Cruz (UCSC). Given the limited number of matched normal tissues in TCGA, we supplemented the normal control group by retrieving RNA-seq data of normal cervical tissues from the Genotype-Tissue Expression (GTEx) project (gtex_RSEM_gene_fpkm). After excluding tumor samples with missing survival information (N = 13), 291 tumor tissue samples were retained for the analysis of the prognostic risk model (see patient enrollment flowchart in Figure S1). The combined normal control cohort comprised 13 samples: 3 adjacent normal tissues from TCGA-CESC and 10 normal cervical tissues from GTEx. All expression data were log₂-transformed (log₂[fpkm + 1]). Given the technical variations inherent to multi-platform data integration, we performed quantile normalization using the “normalizeBetweenArrays” function in the R package “limma”. This approach harmonizes the distribution of expression values between TCGA and GTEx samples, mitigating platform-specific biases. Gene annotation was performed on the RNA-seq data of all samples (304 in total) using the RNA annotation file in GENCODE (https://www.gencodegenes.org/), resulting in an expression matrix of 20,388 mRNAs.

Furthermore, the GSE6213 dataset was downloaded from the Gene Expression Omnibus (GEO) database, including cervical cancer tumor tissues from different patients collected before (n = 18) and during (n = 10) radiotherapy (unpaired). Samples were collected before any treatment and one week after radiotherapy initiation, respectively. The probes were annotated to the corresponding genes using the GPL96 platform files. For multiple probes matching the same gene, the mean expression value was taken as the representative value for that gene. In addition, 332 LRGs were included in this study based on a previously published report [19], which defined LRGs as genes encoding lactylases (EP300), deacylases (HDAC1-3, SIRT1-3), and proteins identified as lactylated by mass spectrometry (327 proteins).

Identification of key genes

The differentially expressed genes (DEGs) were identified between normal and tumor tissue samples using the R package “limma”, with the screening criteria: |log₂ (Foldchange)| ≥1 and Benjamini-Hochberg adjusted P value < 0.05 [20]. Differentially expressed LRGs (DE-LRGs) were obtained by overlapping DEGs and LRGs. Subsequently, univariate Cox regression analysis was conducted based on DE-LRGs to identify prognosis-related genes that were significantly associated with the patients’ overall survival (P < 0.05). Their expression levels in tumor tissues before and during radiotherapy were analyzed based on the GSE6213 dataset. Among these genes, those with reversed expression trends during radiotherapy were identified as key genes, whereas those showing no significant differences or failing to exhibit this specific reversal pattern were excluded from further analysis.

Expression validation of key genes

The GSE9750 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE9750) dataset was downloaded from the GEO database for expression validation of key genes, including 24 normal cervical and 33 cervical cancer tissue samples. To validate the protein expression levels of key genes, immunohistochemistry (IHC) data of tumor and normal tissues were obtained from the Human Protein Atlas (HPA) (https://www.proteinatlas.org/) [21].

Additionally, we validated the expression levels of key genes in 13 normal cervical and 15 cervical cancer tissue samples using real-time quantitative polymerase chain reaction (RT-qPCR). All tissue samples were collected from patients who underwent surgical resection at Xuzhou Central Hospital between March 2020 and July 2025. Eligible patients met the following criteria: newly diagnosed and histopathologically confirmed cervical cancer; age 18–70 years; no prior radiotherapy, chemotherapy, immunotherapy, or molecular targeted therapy before diagnosis; no other malignancies or autoimmune diseases. All enrolled patients were treatment‑naïve. The patient enrollment flowchart was shown in Figure S1, and the clinical characteristics were summarized in Table S1. The protocol was approved by the Ethics Review Committee of Xuzhou Central Hospital (Approval No.: XZXY-LK-20250415-0043), and all participants provided written informed consent. Total RNA was extracted from the tissues using the TRIzol method, and RNA concentration in each sample was measured using a NanoVue Plus microvolume spectrophotometer (BIOCHROM LTD, USA). The mRNA was reverse-transcribed using FastKing cDNA First Strand Synthesis Kit (Cat. No.: KR116, TIANGEN, Beijing, China) to synthesize cDNA. Subsequently, qRT-PCR was conducted on the Gene-9660 fluorescence quantitative PCR analyzer (BIOER, Hangzhou, China) using SuperReal PreMix Plus (SYBR Green) Kit (Cat. No.: FP205, TIANGEN, Beijing, China) with the resulting cDNA. GAPDH and ACTB served as internal controls and were used jointly for data normalization by calculating the mean of their Ct values. The mRNA expression levels were quantified with the 2-ΔΔCT method, and statistical analysis was performed with the T-test in GraphPad Prism software (version 10.4.0), with all primer sequences detailed in Table S2.

Construction and validation of the prognostic risk model

LASSO regression analysis was performed using the R package “glmnet” [22] based on key genes to construct a prognostic risk model. The model was implemented with family = “cox” and a maximum number of iterations (maxit = 10000). A 10-fold cross-validation was applied to select the optimal penalty parameter (lambda.min). To ensure reproducibility, all samples were randomly divided into training and test cohorts at a ratio of 1:1 with a fixed random seed (set.seed(2024)). Risk scores were calculated based on gene expression levels and regression coefficients. The formula was as follows: Risk score = ∑ [Exp (gene) × Coef (gene)], where Exp (gene) represented the expression level of the specific gene, and Coef (gene) represented its coefficient in the LASSO regression analysis. Subsequently, the optimal cut-off value for risk stratification was determined using the surv_cutpoint function from the R package “survminer”, with minprop = 0.3 to ensure at least 30% of patients in each risk group. This cutoff, derived exclusively from the training cohort, was then applied as a fixed threshold to classify patients in the training, testing, and entire cohorts into distinct risk groups. Kaplan-Meier curve analysis was performed using the R packages “survival” and “survminer”, with the log-rank test for comparing overall survival between risk groups (P < 0.05 considered significant). Receiver operating characteristic (ROC) curves were generated using the R package “timeROC” to validate the predictive accuracy of the risk model at 1, 2, and 3 years. The area under the curve (AUC) was reported with 95% confidence intervals (CI). Principal component analysis (PCA) was performed using the R package “FactoMineR” to observe the distribution pattern of samples in the high- and low-risk groups. Additionally, to assess the stability and reliability of our risk model, we performed internal validation using 10-fold cross-validation in the TCGA cohort. We also compared it with three recently published cervical cancer prognostic models. These models were constructed based on immunogenic cell death-related genes (Model 1) [23], pyroptosis-related genes (Model 2) [24], and mitochondria/ferroptosis-related genes (Model 3) [25], respectively. All models were evaluated uniformly on the training cohort, with the 3‑year AUC for survival prediction selected as the unified performance metric.

DEGs and functional enrichment analysis

DEGs between high- and low-risk groups were screened using the R package “limma”. To screen for more available candidate genes for subsequent functional analysis, the screening criteria were relaxed: |log₂ (Foldchange)| ≥0.5 and P < 0.05 [20]. GO function and KEGG pathway analysis were performed using the DAVID (https://david.ncifcrf.gov/) [26, 27] database (P value < 0.05) [20].

Analysis of immune cell infiltration

The infiltration level of immune cells between the low- and high-risk groups was assessed using the single-sample gene set enrichment analysis (ssGSEA) algorithm and statistically analyzed by the rank-sum test. The correlation between risk scores and differentially infiltrating immune cells was assessed using Pearson correlation analysis, with P < 0.05 considered statistically significant [25].

Analysis of tumor mutational burden (TMB)

The somatic mutation data were downloaded from the TCGA database. The R package “maftools” was used to analyze the gene mutation characteristics of the high- and low-risk groups. The TMB of each sample was calculated and compared between the two risk groups, with P < 0.05 considered statistically significant [28].

Development of a nomogram

To explore the independence of risk scores and clinicopathologic features in prognosis assessment, we conducted univariate and multivariate Cox regression analyses using the R “coxph” function. Consistent with the default setting, we employed a complete case analysis, excluding samples with missing values for any covariate. Subsequently, we constructed a nomogram model using the R package “rms”, and the risk score and other clinicopathologic features were included as predictors. The sum of each predictor’s corresponding scores was calculated to predict the survival probability of patients at 1-, 2-, and 3-years. Furthermore, we evaluated the nomogram’s predictive accuracy using a calibration curve and assessed its clinical utility via decision curve analysis (DCA) with the R package “ggDCA”.

Results

Identification of key genes

After data normalization, the expression distributions of TCGA and GTEx samples were substantially homogenized (Figure S2A), allowing for integrated analysis. Using the RNA-seq data integrated from the TCGA and GTEx databases, a total of 4,747 DEGs were identified between tumor and normal tissues. Among them, 2,356 genes were significantly up-regulated, and 2,391 genes were down-regulated in the tumors (Table S3). These DEGs were further intersected with LRGs, and 100 DE-LRGs were screened (Fig. 1A). Subsequently, 22 prognosis-related DE-LRGs were screened through univariate Cox regression analysis (Fig. 1B). Compared with normal tissues, 11 genes were up-regulated and 11 genes were down-regulated in tumor tissues (Fig. 1C). Among them, DDX39B, NOP2, RFC4, and STMN1 were down-regulated during radiotherapy compared to before radiotherapy (Fig. 1D). Notably, RFC4 and STMN1 exhibited opposing expression trends when comparing during vs. before radiotherapy to those observed in tumor vs. normal tissues. Therefore, RFC4 and STMN1 were identified as key genes.

Fig. 1.

Identification of key genes for radiotherapy prognosis associated with lactylation. A Venn diagram for differentially expressed genes and lactylation-related genes. B Forest plot of univariate Cox regression analysis of 22 prognostic-related genes. HR, Hazard Ratio; CI, Confidence Interval. C Box plots of the expression of prognostic-related genes between tumors and normal tissues. (D) Box plots of the expression of prognostic-related genes in tumor tissues before and during radiotherapy

Validation of RFC4 and STMN1 expression

In the GSE9750 dataset, RFC4 and STMN1 were significantly overexpressed in tumor tissues, consistent with the results in TCGA and GTEx datasets (Fig. 2A). IHC results from the HPA database confirmed this finding at the protein level, showing that RFC4 and STMN1 had higher protein expression levels in tumor tissues than in normal tissues (Fig. 2B). Finally, RT-qPCR experiments in our independent sample cohort again validated the upregulation pattern of these two genes in tumors at the mRNA level (Fig. 2C).

Fig. 2.

Expression validation of key genes. A Box plots of RFC4 and STMN1 expression in the GSE9750 dataset. B IHC results of RFC4 and STMN1 based on the HPA database. C RT-qPCR results of RFC4 and STMN1. ***P < 0.001, ****P < 0.0001

Construction and verification of the prognostic risk model

Although only two genes (RFC4 and STMN1) remained after multi-step filtering, LASSO regression was still performed for coefficient shrinkage and overfitting control, rather than for feature selection. Both genes received non-zero coefficients under the optimal λ (Fig. 3A, B), confirming their joint contribution to the risk model. The final coefficients, hazard ratios (HR), 95% CI, and p-values of the two genes are presented in Table S4. The risk score for each tumor sample was calculated using the formula: risk score = (-0.3542×RFC4) + (-0.2751×STMN1). The proportional hazards assumption was tested using Schoenfeld residuals. No significant violation was observed for either RFC4 (P = 0.58) or STMN1 (P = 0.41), confirming the model’s appropriateness (Figure S2B). Additionally, in the entire cohort of 291 patients, 72 events (deaths) were observed, yielding an events-per-variable ratio of 36 for the two-gene model, which was well above the recommended threshold of 10 for the construction of prognostic models [29].

Fig. 3.

Construction of the prognostic risk model for radiotherapy. A Coefficient path plot for LASSO Cox regression analysis. B Selection of the optimal parameter (lambda) in the LASSO model. (C) Distribution of risk scores and survival status in the training cohort. D Kaplan-Meier curve analysis of high- and low-risk groups in the training cohort. E ROC curve analysis of 1-year, 2-year, and 3-year survival in the training cohort. F PCA plot of patients in the high- and low-risk groups. G Boxplot illustrating the distribution of 1-year, 2-year, and 3-year AUC values from 10-fold cross-validation. Red diamonds represent the mean AUC value at each time point. H Comparison of the prognostic models in our study with those in previous studies

Subsequently, the 291 cervical cancer patients (entire cohort) were randomly assigned in a 1:1 ratio to a training cohort (N = 145) and a testing cohort (N = 146). In the training cohort, patients were further divided into the high- and the low-risk groups according to the optimal cut-off value (median risk score = 0.09088121), with the high-risk group having more deaths of patients (Fig. 3C). Kaplan-Meier analysis showed that the low-risk group exhibited a longer survival time than the high-risk group (P = 0.00084, Fig. 3D). ROC curve analysis demonstrated that the AUC of the risk model predicting 1-year, 2-year, and 3-year survival was 0.602 (95% CI: 0.476 − 0.732), 0.714 (95% CI: 0.573 − 0.848), and 0.717 (95% CI: 0.588 − 0.852), respectively (Fig. 3E), suggesting acceptable discriminatory ability for survival prediction. PCA results showed that the samples of the high- and low-risk groups in the training cohort present two different distribution patterns (Fig. 3F). The results of the 10-fold cross-validation showed that the AUC value distributions of our risk model at 1-year, 2-year, and 3-year were 0.557 ± 0.214 (95% CI: 0.156–0.785), 0.638 ± 0.178 (95% CI: 0.369–0.913), and 0.603 ± 0.187 (95% CI: 0.316–0.869), respectively (Fig. 3G). This suggested that our model possessed favorable robustness without obvious overfitting, thereby supporting the reliability of the risk stratification. The comparative analysis with previously published prognostic models found that our prognostic risk model had a higher AUC value (our model: 0.717; Model 1: 0.702; Model 2: 0.658; Model 3: 0.556), suggesting superior prognostic performance (Fig. 3H). Furthermore, the results in the test cohort (Figure S3A–D) and the entire cohort (Figure S3E–H) were consistent with the training cohort, further validating the stability and generalizability of the risk model.

DEGs and functional enrichment analysis in high- and low-risk groups

To explore the potential biological characteristics behind risk stratification, we further analyzed the DEGs between the high- and low-risk groups and performed functional and pathway enrichment analyses. A total of 391 DEGs were identified, with 202 up- and 189 down-regulated genes in the high-risk group (Table S3). GO functional enrichment analysis showed that these DEGs are mainly enriched in extracellular related components (Fig. 4B); participating in biological processes related to immunity and angiogenesis, such as “neutrophil chemoaxis”, “inflammatory response”, and “angiogenesis” (Fig. 4A); affecting molecular functions related to immunity and substance binding, such as “chemokine activity”, “cytokine activity” “protein binding” (Fig. 4C). KEGG pathway enrichment analysis showed that these DEGs were enriched in immune- and metabolism-related pathways, such as “IL-17 signaling pathway”, “TNF signaling pathway”, “Cytokine-cytokine receptor interaction”, “Chemokine signaling pathway”, and “Glycolysis/Gluconeogenesis” (Fig. 5). Overall, the potential biological differences between the high- and low-risk groups were mainly related to immune, angiogenic, and metabolic regulatory processes.

Fig. 4.

GO enrichment analysis of DEGs. A–C Bubble plots showing GO enrichment terms categorized into biological processes (A), cellular components (B), and molecular functions (C). The x-axis represents the gene ratio, the y-axis lists the GO terms, the color of each bubble indicates the significance of enrichment [-log₁₀(P value)], and the size of each bubble corresponds to the number of genes enriched in the term

Fig. 5.

Bubble plot of the KEGG pathway enrichment. Bubble size indicates the number of genes enriched in the term, and bubble color represents the P-value. The lines connecting genes to pathways indicate that the genes are involved in this pathway. Gene. Ratio represents the number of DEGs enriched in the pathway as a proportion of the total number of DEGs

Characteristics of immune cell infiltration in high- and low-risk groups

Immune cell infiltration analysis showed significant differences in 11 immune cell subgroups between the high- and low-risk groups. Specifically, activated B cells, CD56 bright natural killer cells, eosinophils, immature dendritic cells, natural killer T cells, T follicular helper cells, type 17 T helper cells, and type 2 T helper cells showed higher infiltration levels in the low-risk group, while MDSCs, neutrophils, and regulatory T cells exhibited relatively higher infiltration levels in the high-risk group (Fig. 6A). Further correlation analysis showed that the risk score was significantly negatively correlated with the infiltration levels of eosinophils, immature dendritic cells, type 17 T helper cells, type 2 T helper cells, and CD56 bright natural killer cells, while significantly positively correlated with MDSCs, neutrophils, and regulatory T cells (Fig. 6B). These distinct immune profiles between risk groups suggested that our risk model effectively captured the underlying heterogeneity of the cervical cancer immune microenvironment.

Fig. 6.

Immune feature assessment between high- and low-risk groups. A Box plot showing the difference in immune cell infiltration levels between high- and low-risk groups. B Stick plot of the correlation between risk score and differential immune cells. The horizontal axis is the correlation coefficient, where > 0 represents a positive correlation and < 0 represents a negative correlation. The vertical axis is the type of immune cells. The circle size reflects the correlation coefficient’s absolute value, while its color shows the correlation’s P value. ns, not significant, *P < 0.05, **P < 0.01, ****P < 0.0001

TMB analysis in high- and low-risk groups

Given the significant differences in immune cell infiltration between the high- and low-risk groups, coupled with the known association between genomic instability and immune dysregulation and worsened disease prognosis [30], we further investigated the mutation profiles of the two groups. The waterfall plot of somatic mutation distribution showed that TTN and PIK3CA were the most frequently mutated genes in both groups, with mutation rates of 31% and 28% in the high-risk group, and 28% and 26% in the low-risk group, respectively (Fig. 7A, B). Further analysis revealed that a trend toward a higher TMB score was observed in the high-risk group compared to the low-risk group, although this difference was not statistically significant (P = 0.23, Fig. 7C).

Fig. 7.

TMB analysis in high- and low-risk groups. A, B Waterfall plot of somatic mutations in the high- (A) and low- (B) risk groups. C Violin plots of TMB score between high- and low-risk groups

Development of a prognostic nomogram integrating risk score and clinicopathological features

To clarify the independent prognostic value of the risk score and establish a clinically applicable model, we conducted an integrated analysis combining it with traditional clinical parameters. Univariate Cox analysis showed that risk score (HR [95% CI] = 2.45 [1.53–3.90]; P = 0.0002), stage (HR [95% CI] = 1.48 [1.20–1.84]; P = 0.0004), T (HR [95% CI] = 1.83 [1.37–2.44]; P < 0.0001), N (HR [95% CI] = 2.56 [1.30–5.03]; P = 0.0064), and M (HR [95% CI] = 2.55 [1.20–10.55]; P = 0.0225), were all significantly associated with poor prognosis (Fig. 8A). Multivariate Cox analysis showed that risk score (HR [95% CI] = 3.77 [1.38–10.30]; P = 0.01), stage (HR [95% CI] = 0.37 [0.16–0.89]; P = 0.026), and T (HR [95% CI] = 2.64 [1.23–5.66]; P = 0.013) could serve as independent prognostic predictors (Fig. 8B). Subsequently, nomogram models were constructed based on these independent prognostic predictors to predict the 1-year, 2-year, and 3-year survival rates of patients (Fig. 8C). The calibration curves demonstrated that the nomogram model had good prognostic predictive ability (Fig. 8D). DCA indicated that the nomogram had a better effect on patients’ clinical benefits (Fig. 8E). ROC curve analysis showed that the AUC values of the nomogram achieved AUC values of 0.787 (95% CI: 0.590 − 0.872) at 1-year, 0.787 (95% CI: 0.539 − 0.792) at 2-year, and 0.787 (95% CI: 0.504 − 0.726) at 3-year. These values were superior to those of the stage, T, and risk score alone (Fig. 6F), demonstrating that the accuracy of the nomogram in predicting survival rate was relatively high.

Fig. 8.

Development of a prognostic nomogram. A, B Forest plots of univariate (A) and multivariate (B) Cox regression analysis of risk score and clinicopathological features. C Nomogram predicting the 1-year, 2-year, and 3-year survival probability based on risk score and clinicopathological features. D Calibration curves of the nomogram prediction model. E Decision curve analysis of the nomogram prediction model. F ROC curves of the nomogram, risk score, stage, and T at 1-year, 2-year, and 3-year

Discussion

The prognosis assessment of cervical cancer radiotherapy remains a clinical challenge to date [5]. Although imaging-based techniques and traditional molecular markers (such as HPV typing, PD-L1 expression) can be used to predict treatment efficacy, their accuracy is limited by tumor heterogeneity [30–32]. In recent years, protein PTMs represented by lactylation have shown important roles in tumor biological regulation and treatment [33]. However, the systematic study of the LRG signature linked to cervical cancer radiotherapy has been underexplored. In this study, we constructed and internally validated a prognostic model for cervical cancer based on the LRG signature (RFC4 and STMN1), which provided valuable insights for risk stratification. The distinct tumor biological features and survival outcomes between high- and low-risk groups further supported the model’s reliability. In brief, our findings suggested that the risk model constructed based on RFC4 and STMN1 may have certain prognostic value in cervical cancer, and these two genes may serve as potential prognostic biomarkers.

RFC4, known as replication factor C (activator 1) subunit 4, played a pivotal role in DNA replication and repair, and was widely recognized as an oncogene in human cancers [34]. Existing studies have shown that its expression level is a key indicator for predicting the radiotherapy responsiveness in patients with colorectal cancer, breast cancer, and esophageal cancer [35–37]. In cervical cancer, RFC4 was reportedly upregulated and correlated with copy number variants [38]. Knockdown of RFC4 could induce cell cycle arrest and suppress the proliferation of cervical cancer cells [39]. Its high expression was also a prognostic biomarker, and models incorporating RFC4 could enhance the accuracy of prognosis prediction [20, 40, 41]. STMN1 (Stathmin 1), encoding the protein involved in microtubule polymerization, has been reported as a cancer biomarker [42] and validated to associate with chemoradiation resistance [43]. Wang et al. found that STMN1 was significantly associated with a reduced overall survival time in cervical cancer patients [44]. Mechanistically, STMN1 inhibition not only suppressed the proliferation, invasion, and migration of cervical cancer cells but also enhanced their therapeutic sensitivity [45, 46]. Similar to previous studies, our study found that RFC4 and STMN1 were up-regulated in the tumor tissue and down-regulated during radiotherapy, which indicated that two key genes may be associated with cervical cancer progression and radiotherapy prognosis. Notably, no research has reported a regulatory relationship between RFC4 and STMN1 to date. Both were up-regulated in cervical cancer, with RFC4 controlling DNA replication/repair and STMN1 regulating microtubule dynamics, suggesting they might act in parallel pro-tumorigenic pathways through distinct upstream mechanisms. Overall, the cervical cancer prognosis risk model constructed based on the LRG signature has potential reliability.

Cervical cancer patients were divided into high- and low-risk groups based on the optimal cutoff value of risk score, and significant differences in immune, angiogenic, and metabolic profiles were observed between the two groups. Chemokines, as core regulators of tumor immunity, were involved in regulating immune cell infiltration, inflammatory response, and tumor cell behavior through CXC family members in the cervical cancer microenvironment [47, 48]. Prior experimental studies showed that CXC-chemokine-ligand-10 gene therapy can not only inhibit the growth of cervical tumors by regulating microvessel formation [49], but also significantly enhance the radiotherapy efficacy in HeLa cells [50]. Notably, angiogenesis was a key driver for cervical cancer development, metastasis, and treatment response [51–53]. A preclinical study demonstrated that anti-angiogenic agents can inhibit the epithelial-mesenchymal transition and enhance treatment sensitivity in cervical cancer cells [54]. Furthermore, metabolic reprogramming, as one of the important features of malignant tumors, was typically manifested in cervical cancer as enhanced glycolysis, disturbed amino acid metabolism, and lipid accumulation, collectively driving tumor progression and leading to radioresistance [55, 56]. Previous experimental evidence indicated that inhibiting the glycolytic pathway restored the radiosensitivity of cervical cancer cells [57]. In the present study, we found that DEGs between the high- and low-risk groups differed in the functional/pathway enrichment, such as “inflammatory response,” “angiogenesis,” “Cytokine-chemokine receptor binding”, and “Glycolysis/Gluconeogenesis”. These observations suggest potential associations between these functions/pathways and cervical cancer radiotherapy response, and targeting them might be a promising strategy for improving treatment prognosis.

Immune cell infiltration and TMB are two key factors shaping the cervical cancer tumor microenvironment, and their interaction profoundly influences tumor progression and response to radiotherapy [58, 59]. The risk model developed in this study based on RFC4/STMN1 provided important insights for distinguishing between high- and low-risk patients from these two perspectives. The high-risk group showed elevated infiltration of MDSCs, neutrophils, and regulatory T cells. This infiltration pattern was consistent with the prior studies linking the infiltration of these immunosuppressive cells to poor prognosis and chemoradiation resistance [60–62]. Concurrently, this group showed a trend toward higher TMB scores (although not statistically significant, P = 0.23), and had the highest mutation frequencies for TTN and PIK3CA (31% and 28%, respectively). It was reported that PIK3CA mutations activated the PI3K-Akt pathway and were associated with radiotherapy resistance in cervical cancer [63, 64]. Although TTN has not been reported in cervical cancer, its mutation has been shown to affect the scores of metabolism-related genetic variables in gynecological tumors [65]. The coexistence of this “immunosuppressive microenvironment” and “oncogene mutations” may constitute a vicious cycle: genomic instability may shape an immunosuppressive microenvironment by recruiting immunosuppressive cells, which in turn can help genome-unstable cells evade immune clearance. This “immune-genomic” vicious cycle provides a possible biological explanation for the poor prognosis in high-risk patients.

For the low-risk group, we found high infiltration levels of eosinophils, immature dendritic cells, and type 17 T helper cells. Studies showed that cervical cancer patients with high eosinophil levels had a higher survival rate [66, 67], and radiotherapy-induced eosinophilia was also associated with improved patient prognosis [68]. Kim et al. demonstrated that immature dendritic cells were injected into the tumor site after radiotherapy, which induced effective anti-tumor immunity [69]. Type 17 T helper cells (Th17) promoted antitumor responses by facilitating the recruitment of effector Th1 cells to the tumor [70]. This suggested that patients in the low-risk group may have a more active immune microenvironment, which could be associated with a favorable prognosis. Furthermore, this group exhibited lower TMB levels, with TTN and PIK3CA mutation frequencies of 28% and 26%, respectively, which were slightly lower than those in the high-risk group. Based on the aforementioned immune and genetic mutation profiles, it could be speculated that high- and low-risk patients may respond differently to current treatments. The high-risk group, which was enriched with immunosuppressive cells and higher mutations associated with radiotherapy resistance, might collectively contribute to a phenotype resistant to chemoradiotherapy, while the low-risk group, characterized by an active anti-tumor immune microenvironment, might be more sensitive to chemoradiotherapy. Although our study lacked direct data on treatment response, the current inferences preliminarily suggested the potential value of this risk model in predicting differential treatment benefits. Overall, the differences in immune characteristics and gene mutation between the two groups provided a theoretical basis for understanding their differing clinical outcomes and offered a new hypothesis for future exploration of biomarkers for radiotherapy response in cervical cancer.

In the current study, although the risk model included only two genes, RFC4 and STMN1, it was constructed using a rigorous data-driven process. It demonstrated the ability to significantly distinguish patient prognosis in both the training and testing sets. Significant differences between the high- and low-risk groups in immune cell infiltration, TMB, and functional/pathway enrichment suggested that the model effectively captured key signals associated with tumor heterogeneity. Compared with the three previously published models (all based on 3-gene signatures), this model exhibited superior predictive performance, further validating its effectiveness. Furthermore, the nomogram based on risk scores and clinicopathologic features accurately predicted the survival prognosis of cervical cancer patients at 1-year, 2-year, and 3-year. The calibration curves showed that the nomogram predicted overall survival with high predictability. Overall, this RFC4/STMN1-based risk model provided a preliminary theoretical reference for risk stratification and prognostic assessment in cervical cancer, while the differential pathways and frequently mutated genes identified in this study may offer insights for future mechanistic and translational studies.

However, several limitations of this study warrant consideration. Firstly, given the reliance on bioinformatics analysis and retrospective data without an independent external validation cohort, the generalizability and practicability of this model warrant further validation in real-world clinical data, prospective studies, or multicenter external cohorts. Secondly, quantile normalization using normalizeBetweenArrays aligns expression distributions across TCGA and GTEx but does not explicitly model batch effects. Residual platform‑related biases may persist and could affect results. Without correction, systematic biases would dominate biological signals, compromising model generalizability. We acknowledge these limitations and have validated key findings using an independent RT‑qPCR cohort (N = 15), which confirmed the expression patterns of RFC4 and STMN1. Thirdly, while RFC4 and STMN1 originated from a reported LRG set and their prognostic value was confirmed in this study, direct experimental validation was needed to determine whether they were regulated by or functionally dependent on lactylation processes, and how they mediate radiotherapy response in cervical cancer. Fourthly, although our analysis suggests that RFC4 and STMN1 may be involved in the regulation of the cell cycle, proliferation, immunity, angiogenesis, and metabolism, whether these genes causally drive such effects or merely correlate with the observed phenotypes remains to be determined. Finally, due to data availability constraints in the TCGA dataset, potential confounders, including HPV status, treatment heterogeneity, and tumor subtype, were not accounted for in this study, which may affect the generalizability of our findings. Therefore, further research will design large-scale multi-center clinical cohort studies that incorporate these clinical variables to verify the model’s translational potential. ChIP-qPCR should be used to detect histone lactylation marks at RFC4/STMN1 promoters, or functional modulation of lactate metabolism (e.g., LDHA inhibition or lactate supplementation) should be performed to verify lactylation-dependent regulation of these genes. Concurrently, loss- and gain-of-function experiments (e.g., siRNA knockdown or CRISPR-Cas9 editing) in cervical cancer models are warranted to elucidate the causal roles of RFC4 and STMN1.

Conclusion

In summary, this study constructed an exploratory and internally validated prognostic model for cervical cancer based on the LRG signature, namely RFC4 and STMN1, which were linked to radiotherapy-associated transcriptional changes. Patients were divided into high- and low-risk groups according to the optimal cutoff value of the risk score, with preliminary evidence suggesting differential survival outcomes between groups. Further studies revealed potential differences in function/pathway, tumor immune status, and TMB between them. These findings suggested a preliminary value for the prognosis assessment and risk stratification of cervical cancer patients, warranting further investigation to explore their utility in optimizing radiotherapy strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Y.T., L.N., Y.Z., and H.H. conceived and designed this study; Y.T. and X.Y. participated in data collection and collation; L.N., N.L., and X.Y. participated in data analysis, interpretation, and visualization; H.H., N.L., and Y.L. collected the clinical samples and related data; Y.Z. and H.H. provided administrative support; Y.T., L.N., Y.Z., and H.H. wrote and revised the manuscript. All authors read and approved the final manuscript.

Funding

This research was funded by the 2025 Annual Pengcheng Talent Program Project of Xuzhou Municipal Health Commission (No. 2025TD08).

Data availability

All data used and/or analyzed in this study were obtained from public databases and are included in this published article. The source code (R script) required to reproduce the analysis is available at File S1.

Code availability

All data used and/or analyzed in this study were obtained from public databases and are included in this published article. The source code (R script) required to reproduce the analysis is available at File S1.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Review Committee of Xuzhou Central Hospital (Approval No.: XZXY-LK-20250415-0043). All procedures involving human participants were performed in accordance with the ethical standards of the aforementioned committee and the Declaration of Helsinki. Written informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.