DNA damage response signature-based prognostic genes for intrahepatic cholangiocarcinoma: a combined analysis of machine learning and biological experiments

Abstract

Background

Intrahepatic cholangiocarcinoma (ICC) is a highly aggressive subtype of primary liver cancer with insidious onset, early metastasis, and poor prognosis. DNA damage response (DDR) dysfunction is linked to ICC tumorigenesis, progression, and drug resistance. However, the detailed mechanisms remain underexplored.

Methods

DDR activity was quantified by ssGSEA across 6 independent ICC cohorts. DDR risk model was analyzed using 10 machine learning algorithms to construct 101 combined models, and the optimal Lasso + RSF model was adopted by c-index. The core gene SFN was selected for experimental validation. Single‑cell RNA sequencing and spatial transcriptomics were used to explore cellular localization and tissue distribution of signature genes. In vitro assays were performed in ICC cell lines after SFN knockdown to assess cell proliferation and chemosensitivity.

Results

The DDR risk model showed stable prognostic performance in training and external validation cohorts. High‑risk patients exhibited poor prognosis, immunosuppressive microenvironment and reduced sensitivity to multiple chemotherapeutic agents. SFN was specifically enriched in malignant cells. Knockdown of SFN significantly suppressed cell proliferation and colony formation, enhanced sensitivity to cisplatin and gemcitabine, and increased γ‑H2AX expression indicating elevated DNA double‑strand breaks.

Conclusions

We constructed a robust DDR‑based prognostic model for ICC that enables accurate risk stratification and prediction of therapeutic response. SFN serves as a critical driver of chemoresistance and a potential therapeutic target. These results offer novel biomarkers and mechanistic insights to facilitate precision medicine for ICC patients.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-026-04355-7.

Introduction

Primary liver cancer is a major global health issue with increasing incidence and mortality, particularly in regions with high rates of chronic liver disease. It mainly includes hepatocellular carcinoma (HCC) and cholangiocarcinoma [1]. Cholangiocarcinoma is a heterogeneous group of cancers arising from the epithelial lining of the intrahepatic, perihilar, and extrahepatic bile ducts. Intrahepatic cholangiocarcinoma (ICC) has been gaining attention because of its increasing prevalence and poor prognosis. The main treatment modalities include surgical resection, chemotherapy, radiation therapy, and targeted therapy [2, 3]. However, drug and radiation resistance remain significant barriers because of genetic mutations, epigenetic modifications, DNA repair response, and oxidative stress response [4, 5]. Therefore, it is imperative to prioritize the identification of biomarkers, and the development of novel therapeutic strategies aimed at overcoming resistance and enhancing patient outcomes.

The DNA damage response (DDR) refers to a series of repair mechanisms that cells activate upon the identification of DNA damage, with the objective of reinstating normal cellular function. These mechanisms include base excision repair (BER), nucleotide excision repair (NER), mismatch repair (MMR), homologous recombination (HR), and nonhomologous end joining (NHEJ) repair [6, 7]. DDR is a critical mechanism for cellular survival and the preservation of genomic integrity, facilitating the prompt repair of DNA damage to prevent the accumulation of mutations and the potential malignant transformation of cells. Deficiencies in the DDR are frequently observed in tumors, where impairments or mutations in genes associated with the DDR can result in the accumulation of DNA damage, heightened genomic instability, and elevated mutation rates, thereby fostering tumorigenesis [8, 9]. Previous studies have shown that the RNA helicase DDX21 is significantly overexpressed in colorectal cancer. High expression of DDX21 increases the frequency of chromosomal exchanges, delays the repair of homologous recombination, increases replication stress, and leads to genomic instability and tumorigenesis [10, 11]. Urea cycle enzyme arginosuccinate synthase (ASS1) plays a critical role in the transcriptional regulation of cell cycle genes that are influenced by p53. The absence of ASS1 may result in DNA damage and facilitate cell cycle progression, thereby potentially contributing to the emergence of cancerous mutations [12, 13]. Additionally, the loss of ARID1A (AT-Rich Interaction Domain 1A) is associated with defects in DNA repair mechanisms. Following DNA damage, ARID1A localizes to sites of DNA double-strand breaks (DSBs) and is instrumental in the organization of chromatin loops by recruiting RAD21 and CTCF to these breaks. The depletion of ARID1A is correlated with increased accumulation of micronuclei, activation of the cGAS-STING signaling pathway, and upregulation of immune regulatory cytokines after radiotherapy [14]. Consequently, critical regulatory elements that target the DDR pathway are significant in tumor therapy. DDR inhibitors have potential application value in cancer treatment, especially in overcoming treatment resistance. PARP inhibitors can prevent the repair of single-strand breaks, which leads to the accumulation of DNA double-strand breaks and are particularly effective in tumors with BRCA1/2 mutations. Additionally, ATM/ATR inhibitors enhance cellular sensitivity to DNA damage by obstructing cell cycle checkpoints and DNA repair pathways, thereby promoting apoptosis [15, 16].

Gemcitabine in combination with cisplatin represents the current standard first-line chemotherapy regimen for treating cholangiocarcinoma and significantly improves patients’ overall survival (OS) and progression-free survival (PFS). Nevertheless, many patients eventually acquire resistance to this treatment, leading to therapeutic failure [3]. Cholangiocarcinoma cells can repair DNA damage through activation of the DDR pathway, contributing to chemoresistance [17, 18]. Previous studies have demonstrated that PARP inhibitors (e.g., Olaparib) increase the efficacy of gemcitabine by inhibiting the base excision repair (BER) pathway [19]. Additionally, ATM/ATR inhibitors increase the sensitivity to cisplatin by disrupting DDR signaling [20, 21]. The integration of DDR-targeted therapies, such as PARP inhibitors and ATM/ATR inhibitors, presents a promising strategy for overcoming chemoradiotherapy resistance in cholangiocarcinoma.

In this study, we aimed to integrate machine learning with biological experiments to identify and validate key prognostic genes within the DDR signature in ICC. By exploring publicly available bulk RNA sequencing, single-cell RNA sequencing and spatial transcriptomics datasets, we identified 9 DDR-related genes associated with patient prognosis. Moreover, based on the expression and efficiency of the 9 DDR-related genes, we constructed a DDR-related risk score model. The model was effective in predicting survival time and prognosis of patients, characterizing immune landscape, and evaluating therapeutic responses. We subsequently validated the functional roles of the core gene, SFN, through in vitro experiments, providing mechanistic insights into their contributions to ICC progression and therapy resistance.

Materials and methods

Data source and collection

The mRNA expression data and clinical information of the ICC patients were downloaded from the NODE (National Omics Data Encyclopedia, https://www.biosino.org/node/home) as the FU-ICCA cohort [22]. Patients without survival data were excluded, and 244 patients were included for further exploration. Moreover, GSE107943, GSE255058, GSE26566, GSE32879, GSE45001, GSE76297 and GSE119336 datasets were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo) via the “GEO query” R package, and TCGA-CHOL was retrieved from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/) as the validation cohort of our model. DNA damage response (DDR) related genes were collected from previous studies and can be found in Table S1 [23].

Evaluation of DDR activity scores in ICC patients

The ssGSEA (single-sample Gene Set Enrichment Analysis) algorithm, implemented via the “GSVA” package, was used to quantify the DDR pathway activity score for each ICC sample. A pre-curated DDR-related gene set was utilized to calculate a normalized enrichment score (NES) for each individual sample, reflecting the relative enrichment level of DDR-related genes. Differences in DDR scores across tumor/normal tissues, TNM stages, and clinical phenotypes (metastasis/vascular invasion) were quantified via non-parametric tests. Prognostic analysis was performed using Kaplan–Meier (KM) survival curves, and all results were visualized with “ggplot2”.

Construction and validation of the DDR signature-based prognostic model

To explore the DDR genes in the prognosis of ICC patients, further analyses were performed using the “Mime1” R package. The FU-ICCA cohort was used as the training set, while the GSE89748, GSE107943 and TCGA-CHOL cohorts served as validation sets. A DDR signature was then constructed with 10 machine learning algorithms: StepCox, Ridge, plsRcox, Lasso, CoxBoost, Enet, GBMs, SVMs, SuperPC, and RSF.

The 10 original algorithms were combined and subquently obtained 101 machine learning algorithm combinations. The expression profiles from all arrays underwent normalization before analysis. Univariate Cox regression analysis was conducted separately in the training cohort, and prognostic genes were screened using the criteria of p-value < 0.1. Then the “ML.Dev.Prog.Sig” function was utilized to compute the concordance index (C-index) across 101 combinations of machine learning algorithms. The algorithm yielding the highest C-index was selected to establish the DDR risk model. Subsequently, patients were stratified into high DDR and low DDR risk subgroups by riskscore with “rs_sur” function. Furthermore, the “ML.Corefeature.Prog.Screen” function was employed to find the core genes in DDR risk model.

To show the distribution of high-risk and low-risk populations, OS time/event, and model gene expression, a risk factor correlation plot was generated. Then, KM analysis, using the “survival” and “survminer” packages, was performed to assess the probability of survival between the two risk groups. Receiver operating characteristic (ROC) curves were generated to predict the accuracy of this model, and area under the curve (AUC) values for 1-, 2-, 3- and 4-year survival were calculated via the “timeROC” package.

Correlations of the DDR-related risk score with different clinical features

KM analysis was further employed to explore high-risk and low-risk groups in different clinical subtypes, including patients with Stages I-II and III-IV, those with or without vascular invasion, regional lymph node metastasis, intrahepatic metastasis, distal metastasis and perineural invasion.

Immune cell infiltration, immune checkpoint and immune function analyses

The R package “CIBERSORT” was used to analyze immune infiltration between different risk groups. The proportions of 22 classical immune cells were explored, and the frequencies of these cells in the high-risk group and the low-risk group were compared. In addition, we carried out a correlation analysis to elucidate the relationships between immune cell infiltration and the model genes used to construct the model. Patient response to immunotherapy is affected by immune checkpoint levels and immune function. Therefore, we compared the expression of immune checkpoint genes and immune functions between the two risk groups. Immune checkpoint genes were identified from a previous study [24]. The 29 immune signatures (gene sets) were obtained from a published study by ssGSEA algorithms [25].

Construction of the nomogram

Univariate and multivariate Cox regression analyses were performed to identify independent prognostic factors. Variables with p < 0.05 in univariate analysis were enrolled into multivariate analysis. A prognostic nomogram was established based on the independent prognostic factors identified by multivariate Cox regression, age, gender, and stage. The concordance index (C-index) was used to evaluate the discrimination of the nomogram. Calibration curves were plotted to assess the consistency between predicted and actual OS at 1, 2, and 3 years. Decision curve analysis (DCA) was conducted to estimate the clinical net benefit of the nomogram.

Predicting drug and therapy sensitivity

To further predict the therapy sensitivity of different DDR risk groups, we calculated the IC50 values of commonly used chemotherapeutic agents in ICC by the “oncoPredict” R package. The GSE255058 dataset comprised 18 patients with advanced intrahepatic cholangiocarcinoma (ICC) and was designed to identify biomarkers associated with differential responses to hepatic arterial infusion chemotherapy combined with lenvatinib and PD-1 inhibitors. We further compared the mRNA expression levels of the model genes between responders and non-responders in this cohort.

Single-cell RNA sequencing data acquisition and processing

The single-cell RNA sequencing (scRNA-seq) dataset GSE138709 of ICC patients was obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). Clinical information was extracted via the “tinyarray” package in R. The raw data files were downloaded and processed to create a Seurat object for downstream analysis. Quality control (QC) metrics, including the number of genes, mitochondrial gene percentage, ribosomal gene percentage, and hemoglobin gene percentage, were calculated to filter out low-quality cells. Cells with extreme values for these metrics were excluded from further analysis. After quality control, the scRNA-seq data were normalized via the “NormalizeData” R function, and highly variable genes were identified via the “FindVariableFeatures” function. Data integration was performed via the “Harmony” algorithm to correct for batch effects. Principal component analysis (PCA) was conducted, followed by t-distributed stochastic neighbor embedding (t-SNE) for dimensionality reduction. Clustering was performed via the “FindNeighbors” and “FindClusters” functions, with a resolution of 0.5.

Cell type annotation was performed via manual methods. For manual annotation, marker genes for each cell type were obtained from the literature [26] and visualized via dot plots. Differential gene expression analysis was conducted via the “FindAllMarkers” R function to identify marker genes for each cell type. Enrichment analysis, including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses, was performed via the “tinyarray” R package to explore the biological functions of the model genes. The “AUCell” algorithm was used to calculate the activity of the model genes to construct the prognostic model in each cell. The cells were classified into high- and low-activity groups based on the AUCell score threshold.

Spatial transcriptomics analyses

The spatial transcriptome dataset was downloaded from the China National Center for Bioinformation database (CNCB) and can be accessed through the project number PRJCA003792 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA000437). We employed a deconvolution analysis technique to evaluate each point’s cell composition on a 10 × Visium slide. We used the scRNA-seq data from the ICC samples for reference. To ensure the reliability of the analysis results, we implemented quality control measures based on the number of expressed genes, the count of unique molecular identifiers (UMI), and the percentage of mitochondrial RNA in each cell. We subsequently constructed a signature scoring system by calculating the top 25 gene expression levels of each cell type and successfully built an enrichment scoring matrix via the “Cottrazm” R package. Finally, using the “SpatialFeaturePlot” function of the “Seurat” package, we successfully annotated different cell types. Using the “AUCell” R package, we analyzed the location of the model genes in ICC tissue. The “SpatialFeaturePlot” R function was used for visualization of the gene expression landscape in each microzone. Spearman correlation analysis was performed between the model genes and the microenvironment components at the spatial transcriptomic level.

Cell culture

The human ICC cell line HuCCT1, RBE and the human embryonic kidney 293 T (HEK293T) cell line were purchased from the National Collection of Authenticated Cell Cultures (Shanghai, China). The cell lines were cultured in DMEM (C11995500BT, Gibco) or RPMI 1640 (C22400500BT, Gibco) supplemented with 10% fetal bovine serum (FBS; 1027–106, Gibco) and 1% penicillin and streptomycin (15,070,063, Gibco) at 37 °C in 5% CO₂.

Plasmids, antibodies and reagents

SFN shRNA was purchased from Tsingke Biotechnology.The detailed sequences of all shRNAs were listed in Table S2. The anti-SFN antibody used for western blotting was purchased from Proteintech (66,251–1-Ig). The anti-Tubulin antibody used for western blotting was purchased from ServiceBio (GB15140-100). The anti-γ-H2AX antibody used for western blotting was purchased from Immunoway (YM8195). Cisplatin (HY-17394) and gemcitabine (HY-17026) were purchased from MedChemExpress.

Western blot, cell proliferation and viability

The knockdown efficiency of SFN and the DDR marker γ-H2AX were detected via western blotting. The detailed methods used for western blotting were described earlier [27, 28]. For the cell proliferation assays, 10 µL of CCK-8 solution (Yeasen, 40203ES80) was added to each well of a 96-well plate after culture for 24 h, 48 h and 72 h. The methods for the colony formation and drug sensitivity assays were the same as those previously reported [29, 30].

Statistical analysis

All the statistical analyses were performed via R (version 4.3.2) and GraphPad Prism 9. Two risk groups were compared via the Wilcoxon signed-rank test and unpaired t-test. For multiple-group comparisons, the Kruskal‒Wallis test and one-way ANOVA were employed. All the statistical tests were two-sided, and P < 0.05 was considered statistically significant.

Results

DDR activation correlates with malignant clinical features in ICC

The data of the FU-ICCA cohort included 255 patients with ICC and 244 patients with survival status and were subsequently selected for the following analysis. The overall design of our study was shown in the graphical abstract . To systematically characterize the clinical relevance of the DDR in ICC, we first explored DDR scores across 6 independent public cohorts (TCGA-CHOL, GSE26566, GSE32879, GSE45001, GSE76297, GSE119336) by ssGSEA, and observed that ICC tumor tissues exhibited significantly higher DDR scores than normal liver tissues in all datasets (p < 0.01 or p < 0.001, Wilcoxon rank-sum test, Fig. 1A–F), confirming a consistent hyperactivated DDR phenotype in ICC. Building on this pan-cohort validation, we further evaluated the prognostic implication of DDR activity in FU-ICCA cohort; Kaplan–Meier (KM) survival analysis revealed that patients with elevated DDR scores presented inferior overall survival relative to those with low scores (HR = 1.98, 95% CI: 1.31–2.95, p = 0.001, Fig. 1G). Moreover, higher DDR scores were observed in patients with distal metastasis, intrahepatic metastasis and vascular invasion (p = 2.6e-03, p = 0.02 and p = 8.2e-03, respectively; Fig. 1H-J). Collectively, these findings demonstrated that intensified DDR signaling is tightly linked to malignant progression and poor clinical outcomes in ICC, supporting its potential value as a prognostic biomarker and therapeutic target.

Fig. 1.

Elevated DDR score predicts poor prognosis and aggressive clinicopathological features in ICC (A-F) Violin plots demonstrate significantly higher DDR scores in ICC tumor tissues compared to normal liver tissues across 6 independent cohorts (TCGA-CHOL, GSE26566, GSE32879, GSE45001, GSE76297, GSE119336). (G) Kaplan–Meier survival analysis of ICC patients (FU-ICCA cohort) stratified by low vs. high DDR scores, showing significantly worse overall survival in the high-score group. (H-J) Box plots illustrating the correlation between DDR scores and clinicopathological characteristics in the FU-ICCA cohort: (H) distal metastasis, (I) intrahepatic metastasis, (J) vascular invasion. *P < 0.05, **P < 0.01, ***P < 0.001 (Wilcoxon rank-sum test for two groups; Kruskal–Wallis test for TNM stage)

Machine learning‑driven development of a DDR prognostic signature for ICC

To establish a prognostic signature for ICC based on DDR genes, we evaluated 101 integrated machine learning models and compared their performance using C-index (Fig. 2A). The top 9 models with highest predictive accuracy were further visualized in Fig. 2B, by which “Lasso + RSF” was determined as the optimal combination. The optimal model exhibited stable and favorable performance in both training and validation across four independent ICC cohorts (Fig. 2C). According to gene selection frequency, we identified the top 8 core prognostic genes (Fig. 2D). We then performed Lasso regression to screen key genes and determine their coefficients for signature construction (Fig. 2E–F). Using the RSF algorithm, we further evaluated variable importance and model tuning to improve prognostic stratification (Fig. 2G–H). Together, these data support the construction of a DDR genes–based prognostic signature for ICC via the optimized Lasso + RSF machine learning framework.

Fig. 2.

Construction and validation of a DDR gene–based prognostic signature via integrative machine learning. (A) Performance comparison of 101 machine learning–based prognostic model combinations assessed by C-index in 4 ICC datasets (FU-ICCA, GSE89748, GSE107943, and TCGA-CHOL). (B) Magnified view of the top 9 best-performing models selected from (A), identifying Lasso + RSF as the optimal combination. (C) Mean C-index of each model in training and validation across 4 independent ICC cohorts: FU-ICCA, GSE89748, GSE107943, and TCGA-CHOL. (D) Top 8 genes ranked by screening frequency across multiple machine learning workflows. (E, F) Lasso regression–based gene screening and coefficient distribution for signature construction. (G, H) RSF model–based variable importance and tuning curves for prognostic stratification

Prognostic validation of the DDR risk model in multiple independent ICC cohorts and stratified subgroup analysis.

To comprehensively validate the prognostic performance of the Lasso + RSF-derived DDR risk signature established in Fig. 2A, we evaluated its predictive efficacy across four independent intrahepatic cholangiocarcinoma (ICC) cohorts: FU-ICCA, GSE89748, GSE107943, and TCGA-CHOL. As shown in Fig. 3A, D, G, J, patients with higher risk scores exhibited significantly shorter survival time and a higher proportion of death events across all four datasets, a pattern further confirmed by the risk score distribution plots (Fig. 3B, E, H, K). Kaplan–Meier survival analysis demonstrated that patients in the high-risk group had substantially worse overall survival than those in the low-risk group in all cohorts, with consistent and statistically significant hazard ratios: FU-ICCA (HR = 27.19, 95% CI: 13.02–56.76, p < 0.001), GSE89748 (HR = 2.53, 95% CI: 1.33–4.82, p = 0.005), GSE107943 (HR = 4.31, 95% CI: 1.47–12.62, p = 0.008), and TCGA-CHOL (HR = 3.02, 95% CI: 1.03–8.86, p = 0.044) (Fig. 3C, F, I, L). To further assess the clinical applicability of the DDR risk signature, we performed stratified survival analysis across key clinicopathological subgroups of ICC patients (Supplementary Fig. 1). Consistent with the overall cohort findings, the high-risk group exhibited significantly inferior overall survival compared to the low-risk group across nearly all tested subgroups, including gender (both p < 0.001, Supplementary Fig. 1A), aged ≥ 65 and < 65 years (both p < 0.001, Supplementary Fig. 1B), early-stage (I-II) and advanced-stage (III-IV) disease (both p < 0.001, Supplementary Fig. 1C), patients with and without distant metastasis (p = 0.008 and p < 0.001, respectively, Supplementary Fig. 1D), intrahepatic metastasis (both p < 0.001, Supplementary Fig. 1E), perineural invasion (both p < 0.001, Supplementary Fig. 1F), regional lymph node metastasis (both p < 0.001, Supplementary Fig. 1G), and vascular invasion (both p < 0.001, Supplementary Fig. 1H).

Fig. 3.

Prognostic performance and risk stratification of the DDR risk model. (A, D, G, J) Scatter plots showing the distribution of survival time and survival status (alive = blue, dead = red) in patients stratified by increasing Lasso + RSF DDR risk score, across four independent ICC cohorts: FU-ICCA, GSE89748, GSE107943, TCGA-CHOL. (B, E, H, K) Risk score distribution plots for patients ranked by increasing risk score in FU-ICCA, GSE89748, GSE107943, TCGA-CHOL. (C, F, I, L) Kaplan–Meier survival curves comparing overall survival between high-risk and low-risk groups in 4 cohorts: FU-ICCA, GSE89748, GSE107943, TCGA-CHOL

Collectively, these results confirm that the model-derived DDR risk signature is a robust and generalizable prognostic biomarker for ICC, enabling accurate risk stratification across multiple independent patient cohorts.

Clinical correlation analysis and independent prognostic value of the DDR risk signature in ICC

We next analyzed the correlation between the DDR risk score and key clinicopathological characteristics in the FU-ICCA cohort, finding that risk scores increased progressively with advancing TNM stage, and were significantly higher in patients with intrahepatic metastasis, distal metastasis, perineural invasion, regional lymph node metastasis, or vascular invasion compared to those without these aggressive features (all p < 0.01, Fig. 4A-F). Time-dependent ROC analysis further confirmed the predictive performance of the signature, with consistently high AUC values for 1-year, 2-year, 3-year, and 4-year overall survival (Fig. 4G-J). These findings indicate that the DDR risk signature is tightly associated with malignant progression in ICC, serving as a generalizable prognostic biomarker that enables accurate risk stratification and survival prediction.

Fig. 4.

Correlation of DDR risk score with clinicopathological features and prognostic predictive performance in ICC (A) The distribution of DDR risk scores across different AJCC tumor stages (I, II, III, IV) in ICC patients. (B-F) Box plots comparing DDR risk scores between ICC patients with and without (B) intrahepatic metastasis, (C) distal metastasis, (D) perineural invasion, (E) regional lymph node metastasis, and (F) vascular invasion. (G-J) Time-dependent ROC curves evaluating the prognostic performance of the DDR risk model for 1-year (G), 2-year (H), 3-year (I), and 4-year (J) overall survival in the training cohort (FU-ICCA) and external validation cohorts (GSE89748, GSE107943, TCGA-CHOL)

To further establish the clinical utility of the DDR risk signature, we performed univariate Cox regression analysis, which revealed that the DDR risk score, TNM stage, intrahepatic metastasis, vascular invasion, regional lymph node metastasis, perineural invasion, tumor size, CA19‑9, CEA, ALT, and GGT were all significantly associated with overall survival in ICC patients (Fig. 5A). Subsequent multivariate Cox regression analysis confirmed that the DDR risk score remained an independent prognostic factor for overall survival (HR = 5.64, 95% CI: 4.22–7.54, P < 0.001) after adjusting for other confounding clinicopathological variables (Fig. 5B). To facilitate the practical application in clinical settings, we constructed a nomogram integrating the DDR risk score, significant independent prognostic factors, as well as age, gender, and TNM stage, to quantitatively predict 1-, 2-, and 3-year overall survival rates for ICC patients (Fig. 5C). Calibration curves demonstrated excellent agreement between the survival probabilities predicted by the nomogram and the actual observed outcomes (Fig. 5D), validating the accuracy of the predictive model. Decision curve analysis further verified that the integrated nomogram yielded greater clinical net benefit than the DDR risk score alone or conventional clinicopathological indicators alone (Fig. 5E). Concordance index comparison additionally confirmed that the nomogram achieved superior predictive accuracy compared with the clinicopathological model alone or the DDR risk score alone (Fig. 5F).

Fig. 5.

Independent prognostic value and clinical utility of the DDR risk signature in ICC (A-B) Univariate and Multivariate Cox regression analysis evaluating the prognostic impact of clinicopathological variables and the DDR risk score in ICC patients. (C) Nomogram integrating the DDR risk score and significant clinical characteristics for predicting 1, 2, and 3-year overall survival in ICC patients. (D) Calibration curves assessing the consistency between nomogram-predicted and observed 1-, 2-, and 3-year overall survival. (E) Decision curve analysis (DCA) evaluating the clinical net benefit of the nomogram, DDR risk score, and conventional clinicopathological factors. (F) Comparison of concordance index (C-index) among the nomogram, DDR risk score, and clinicopathological model, demonstrating superior predictive accuracy of the integrated nomogram. HR, hazard ratio; CI, confidence interval; OS, overall survival

Collectively, these results demonstrate that the DDR risk signature functions as an independent prognostic biomarker in ICC. When incorporated into a clinical nomogram combined with key clinicopathological factors, it significantly improves the precision and clinical applicability of individual survival prediction, providing a valuable tool for personalized clinical management of ICC patients.

High DDR risk score correlates with an immunosuppressive microenvironment and drug resistance in ICC

The tumor microenvironment (TME) plays a critical role in modulating tumor progression, metabolic reprogramming, therapeutic response, and clinical prognosis. Using the CIBERSORT algorithm, we comprehensively profiled the infiltration landscape of 22 immune cell subsets between the low-risk and high-risk groups. Analysis of immune infiltration profiles revealed significant differences in the abundance of multiple immune cell subsets between high- and low-risk groups (Supplementary Fig. 2A). Consistently, the model genes exhibited strong correlations with the infiltration levels of key immune cells, as demonstrated by the correlation heatmap (Supplementary Fig. 2B), suggesting the potential role of these genes in regulating tumor immune microenvironment and anti-tumor immunity.

In cancer therapy, immune checkpoint inhibitors (ICIs), such as PD-1, PD-L1, and CTLA-4, block these checkpoints to increase T-cell activation and antitumor immunity. Therefore, we explored the differences in the mRNA expression of immune checkpoint genes between the low-risk and high-risk groups. There were significant differences in several genes between the two groups (Supplementary Fig. 2C). Among these immune checkpoint-related genes, CD274, CD47, CD70, CD80, CD86, CEACAM1, PDCD1LG2, PVR, TNFRSF18, TNFRSF9, TNFRSF4, TNFSF18, TNFSF4, and TNFSF9 were significantly more highly expressed in the high-risk group. In contrast, BTNL9, CD160, CD28, CD40LG, SIRPA, and VTCN1 were significantly upregulated in the low-risk group.

Immune function analysis involves a comprehensive evaluation of the immune system’s response and capabilities. Based on 27 different immune processes with specific markers, we detected increased levels of increased APC co-inhibition, CCR, DCs, macrophages, para-inflammation, and Tregs in the high-risk group, suggesting an immunosuppressive tumor microenvironment in the high-risk subgroup (Supplementary Fig. 2D).

We further assessed the predictive value of the risk signature for chemotherapeutic response. Notably, high-risk patients showed significantly lower sensitivity (higher IC50 values) to afatinib (p = 1.7e-05), docetaxel (p = 5.3e-06), oxaliplatin (p = 0.038), sorafenib (p = 0.001), rapamycin (p = 1.5e-06), and KRAS(G12C) inhibitor-12 (p = 0.015) compared with low-risk patients (Supplementary Fig. 2E-J). Collectively, these findings indicate that the DDR risk signature is closely linked to the immune landscape and chemosensitivity of ICC, providing a rationale for personalized therapeutic strategies based on risk stratification.

Single-cell transcriptomic analysis reveals the cellular localization and functional characteristics of the DDR risk signature genes in ICC

The single-cell landscape of ICC patients was characterized via the GSE138709 dataset. Quality control metrics, including the number of genes, mitochondrial gene percentage, ribosomal gene percentage, and hemoglobin gene percentage, were used to filter out low-quality cells (Supplementary Fig. 3A). After that, a total of 19 cell clusters were identified via the t-SNE method (Fig. 6A). These clusters were annotated into 10 major cell types, including B cells, cholangiocytes, dendritic cells (DCs), endothelial cells, fibroblasts, hepatocytes, macrophages, malignant cells, NK cells, and T cells (Fig. 6B). The distribution of cell types across samples revealed that tumor samples (ending with "T") and adjacent normal samples (ending with "P") presented distinct cellular compositions (Fig. 6C, D). The marker genes used for annotation were consistent with those reported in the original dataset (Supplementary Fig. 3B).

Fig. 6.

The single-cell landscape of the LIHC patients. (A)The visualization of 19 cell clusters in GSE138709 using the t-SNE method. (B) The cell type annotation of GSE138709. We identified 10 cell types, including B cells, cholangiocytes, dendritic cells (DCs), endothelial cells, fibroblasts, hepatocytes, macrophages, malignant cells, NK cells, and T cells. (C) The number of cells in every cell type. (D) The proportion of cells in every sample. The samples ended with “P” were the adjacent normal samples near the tumor, while the samples ended with “T” were tumor samples. The same number meant the samples were from the same patients. The detailed correspondence between the samples and patients could be viewed from the GEO dataset (https://www.ncbi.nlm.nih.gov/geo/). (E–F) The expression of model genes constructing the prognostic model. The SFN and CDK6 was highly expressed in malignant cells. While the other had relatively higher expression levels in NK cells. (G) The cells are separated into high and low AUC groups using the “AUCell” method. (H) The cells with high AUCell scores were mainly malignant cells and NK cells, which showed that the model genes had a link to the malignant transition and the activation of NK cells

The expression of model genes (SFN, CDK1, PMAIP1, RFC1, CASP3, CCNB2, CDK6, SESN1 and RB1) was analyzed across different cell types. SFN and CDK6 were highly expressed in malignant cells, whereas CDK1, PMAIP1, RFC1 CASP3 and CCNB2 presented relatively high expression levels in NK cells (Fig. 6E). A heatmap of gene expression further confirmed these findings (Fig. 6F). Enrichment analysis revealed that these genes were significantly associated with the p53 signaling pathway (KEGG) and the T cell homeostasis (Supplementary Fig. 3C-D).

The “AUCell” algorithm was used to assess the activity of the model genes in different cell clusters. The cells were divided into high- and low-AUC groups based on a threshold of 0.043 (Fig. 6G, Supplementary Fig. 3E). High AUC scores were predominantly observed in malignant cells and NK cells, suggesting a link between the model genes and malignant transformation or NK cell activation (Fig. 6H). A violin plot further illustrated the distribution of AUC scores across different cell types (Supplementary Fig. 3F).

Spatial transcriptomic analysis uncovers the tissue localization and tumor microenvironment relevance of the DDR prognostic model genes in ICC

To investigate the spatial distribution and functional relevance of our prognostic model genes in ICC, we performed comprehensive spatial transcriptomic analyses on tumor tissue sections from ICC patients. As illustrated in Fig. 7A, the original pathological section of the ICC patient demonstrated typical histological features of cholangiocarcinoma. Subsequent cell type annotation (Fig. 7B) revealed a heterogeneous tumor microenvironment (TME), where each spot represented a microscale tissue region (~ 55 μm in diameter) classified according to its predominant cellular composition. The AUCell scoring analysis (Fig. 7C) revealed significant enrichment of the model genes in the tumor cell regions. Furthermore, Spearman correlation analysis (Fig. 7D) revealed positive associations between the model genes and key TME components, including CD8 + T cells, macrophages, neutrophils, and tumor cells, indicating their involvement in immune modulation and malignant progression. Finally, spatial mapping (Fig. 7E) confirmed the tissue localization of the the model genes (SFN, CDK1, PMAIP1, RFC1, CASP3, CCNB2, CDK6, SESN1 and RB1), with distinct expression hotspots corresponding to tumor-invasive fronts, immune-rich zones, or perivascular regions.

Fig. 7.

The Spatial transcriptomics analyses of the model genes. (A) Original pathological section of the ICC patient. (B) The cell type annotation of the ICC patient. Each dot in the figure represented a tiny tissue region with a diameter of approximately 55 μm, and each micro region is named for the cell type with the largest proportion. (C) The AUCell scoring results of the model genes. (D) The Spearman correlation analyses between the model genes and the microenvironment components. (E) The localization of the model genes in tissues

SFN, a core DDR signature gene, drives oncogenic pathways and predicts prognosis and immunotherapy response in ICC

Having established the “Lasso + RSF”-derived DDR prognostic signature and validated its clinical utility, we next focused on SFN, a top-ranked core gene in the signature (Fig. 2D), to explore its functional role and clinical relevance in ICC. First, we analyzed the correlation between 9 model genes and 14 cancer hallmark pathways, revealing that SFN exhibited significant associations with multiple oncogenic pathways, including cell cycle, DNA damage, EMT, angiogenesis, and apoptosis (Fig. 8A). Subsequent scatter plot analysis confirmed that SFN expression was significantly positively correlated with the activity scores of these pro-tumorigenic pathways, while inversely associated with quiescence and stemness signatures (Fig. 8C). Notably, in the GSE255058 immunotherapy cohort, SFN expression was significantly higher in patients who non-responded to HAIC + Lenvatinib + PD-1 inhibitor combination therapy compared with responders, indicating its potential as a predictive biomarker for immunotherapy efficacy (Fig. 8B). We further validated SFN upregulation in ICC tumor tissues across four independent public cohorts (GSE26566, GSE32879, GSE76297, TCGA-CHOL) relative to normal liver tissues (Fig. 8D-G). Clinically, high SFN expression was strongly associated with poor overall survival in the FU-ICCA cohort (HR = 3.03, 95% CI: 2.02–4.56, p < 0.001, Fig. 8H), and was significantly correlated with advanced TNM stage, regional lymph node metastasis, and perineural invasion (Fig. 8I-L). Collectively, these findings identify SFN as a key functional driver of the DDR signature, linking DDR activation to oncogenic pathway activation, aggressive tumor behavior, and immunotherapy response in ICC.

Fig. 8.

Functional characterization and clinical validation of SFN, a core gene of the DDR prognostic signature in ICC (A) The correlation between key DDR signature genes and 14 cancer hallmark pathway activity scores, highlighting the strong association of SFN with multiple oncogenic pathways. (B) SFN mRNA expression in ICC patients treated with HAIC + Lenvatinib + PD-1 inhibitor, stratified by responder (blue) vs. non-responder (red) from the GSE255058 cohort. (C) The correlation between SFN expression and 14 oncogenic pathway scores (angiogenesis, apoptosis, cell cycle, DNA damage, EMT, etc.) in ICC patients. (D-G) The mRNA expression of SFN in ICC tumor tissues compared with normal liver tissues across four independent cohorts: GSE26566, GSE32879, GSE76297, TCGA-CHOL. (H) Kaplan-Meier survival analysis showing that high SFN expression predicts poor overall survival in the FU-ICCA cohort. (I-M) The association of SFN expression with clinicopathological characteristics in the FU-ICCA cohort: (I) age, (J) TNM stage, (K) regional lymph node metastasis, (L) perineural invasion. p values are calculated using the Wilcoxon rank-sum test or Kruskal-Wallis test. *P < 0.05, **P < 0.01, ***P < 0.001

SFN knockdown inhibits ICC cell proliferation and enhances chemosensitivity via impaired DNA damage repair

To functionally characterize the role of SFN in ICC progression and chemoresistance, we performed loss-of-function experiments in two ICC cell lines, HuCCT1 and RBE, using two independent short hairpin RNAs (shSFN#1 and shSFN#2). Western blot analysis confirmed efficient knockdown of SFN expression in both cell lines (Fig. 9A, H). CCK-8 assays demonstrated that SFN knockdown significantly inhibited the proliferation of HuCCT1 and RBE cells in a time-dependent manner (Fig. 9B, I). Consistent with these findings, colony formation assays revealed that SFN knockdown markedly reduced the number and size of colonies in both cell lines, indicating impaired colony formation ability (Fig. 9C, J). Given the critical role of the DDR in chemotherapy resistance, we next investigated whether SFN knockdown modulates the sensitivity of ICC cells to first-line chemotherapeutic agents, cisplatin and gemcitabine. CCK-8 assays showed that SFN knockdown significantly reduced the half-maximal inhibitory concentration of cisplatin and gemcitabine in HuCCT1 cells, as well as cisplatin in RBE cells, indicating enhanced chemosensitivity (Fig. 9D, F, K). To explore the underlying mechanism, we assessed the expression of γ-H2AX, a well-established marker of DNA double-strand breaks (DSBs). Western blot analysis revealed that SFN knockdown significantly upregulated γ-H2AX expression in HuCCT1 cells treated with cisplatin or gemcitabine, as well as in cisplatin-treated RBE cells (Fig. 9E, G, L). Collectively, these results demonstrate that SFN promotes ICC cell proliferation and chemoresistance, and its knockdown impairs DNA damage repair, leading to the accumulation of DSBs and enhanced chemosensitivity in ICC cells.

Fig. 9.

SFN knockdown suppresses proliferation and enhances chemosensitivity in ICC cells by impairing DNA damage repair (A) Western blot analysis of SFN expression in HuCCT1 cells after transfection with shSFN#1, shSFN#2, or negative control (NC) shRNA. Tubulin served as the loading control. (B) CCK-8 assay showing cell viability of HuCCT1 cells at 0, 24, 48, and 72 h post-transfection. (C) Representative images (left) and quantification (right) of colony formation assays in HuCCT1 cells with SFN knockdown. (D, F) CCK-8 assays evaluating cell viability of HuCCT1 cells treated with increasing concentrations of cisplatin (D) or gemcitabine (F) for 48 h. (E, G) Western blot analysis (left) and quantification (right) of SFN and γ-H2AX expression in HuCCT1 cells treated with 20 μM cisplatin (E) or 20 μM gemcitabine (G) for 48 h. (H) Western blot analysis of SFN expression in RBE cells after SFN knockdown. (I) CCK-8 assay showing cell viability of RBE cells at 0, 24, 48, and 72 h post-transfection. (J) Representative images (left) and quantification (right) of colony formation assays in RBE cells with SFN knockdown. Scale bar, 100 μm. (K) CCK-8 assay evaluating cell viability of RBE cells treated with increasing concentrations of cisplatin for 48 h. (L) Western blot analysis (left) and quantification (right) of SFN and γ-H2AX expression in RBE cells treated with 10 μM cisplatin for 48 h. Data are presented as mean ± SD from three independent experiments. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, two-tailed Student’s t-test or one-way ANOVA followed by Tukey’s post-hoc test

Discussion

ICC is a malignant neoplasm that arises from the epithelial cells of the intrahepatic bile ducts. As the second most prevalent subtype of primary liver cancer, the incidence of ICC has been increasing annually. It is frequently diagnosed at advanced stages, resulting in a poor prognosis [1, 3]. The primary treatment modalities currently employed include surgical resection, chemotherapy, and radiotherapy [2]. Nevertheless, treatment resistance continues to be a significant factor contributing to the unfavorable outcomes observed in patients with ICC [31]. Recent investigations have indicated that impairments in the DDR can lead to the accumulation of DNA damage, thereby facilitating the onset and progression of ICC. Studies have demonstrated that ICC patients exhibiting DDR deficiencies may exhibit heightened sensitivity to specific targeted therapies and immunotherapies [18, 32, 33]. Consequently, therapeutic strategies targeting the DDR system may represent a promising avenue for the personalized treatment of ICC. Nevertheless, the specific regulatory mechanisms by which the DDR pathway modulates ICC progression, prognosis, and therapeutic resistance remain largely unclear.

In the present study, we integrated bulk RNA-seq, single-cell RNA-seq, spatial transcriptomics, machine learning, and in vitro functional experiments to systematically investigate the clinical significance and biological role of DDR-related genes in ICC. We first assessed DDR pathway activity across six independent cohorts using ssGSEA and confirmed that DDR scores were significantly higher in ICC tumor tissues than in normal liver tissues. High DDR activity was significantly associated with advanced TNM stage, distant metastasis, intrahepatic metastasis, vascular invasion, and poor overall survival, indicating that hyperactivated DDR signaling is tightly linked to the malignant progression of ICC. Subsequently, we constructed 101 combined models using 10 machine learning algorithms and selected the optimal “Lasso + RSF” model based on C-index to establish a DDR-related prognostic signature. The signature exhibited robust and stable prognostic performance in the training cohort (FU-ICCA) and multiple external validation cohorts (GSE89748, GSE107943, TCGA-CHOL). The DDR risk score was identified as an independent prognostic factor for ICC patients, and a nomogram integrating the risk score and clinical features was further constructed, which showed superior predictive accuracy and clinical net benefit compared with traditional clinicopathological indicators. These results demonstrate that our DDR-related signature can effectively achieve accurate risk stratification and individualized survival prediction for ICC patients.

The tumor microenvironment (TME), which comprises all noncancerous host cells and non-cellular components, is essential for tumor progression and therapeutic efficiency [34, 35]. To further explore the difference in the TME between the high-risk and low-risk groups, we performed analyses focusing on the immune landscape, immune function, and sensitivity to therapy. Further analysis revealed that the high DDR risk group was characterized by an immunosuppressive tumor microenvironment, with increased infiltration of M0 macrophages and Tregs, upregulated expression of immune checkpoint molecules including CD274 (PD-L1), and enhanced APC co-inhibition, CCR, and para-inflammatory responses. In addition, the high-risk group exhibited higher IC50 values for multiple chemotherapeutic agents (afatinib, docetaxel, oxaliplatin, sorafenib, rapamycin), indicating reduced chemosensitivity. These findings suggest that the DDR signature can reflect the immune landscape and therapeutic sensitivity of ICC, providing a theoretical basis for risk-stratified personalized treatment strategies.

Single-cell RNA-seq analysis showed that SFN and CDK6 were specifically enriched in malignant ICC cells, while other signature genes were mainly distributed in NK cells. Spatial transcriptomics further verified that these model genes were predominantly located in tumor cell regions and were positively correlated with CD8 + T cells, macrophages, and neutrophils. These results reveal the cellular and spatial localization characteristics of DDR-related genes, highlighting their critical roles in tumor malignant progression and tumor-immune crosstalk. Enrichment analysis confirmed that these genes were mainly involved in the p53 signaling pathway, which is consistent with their function in regulating DDR and tumor progression.

As the core gene of the DDR signature, SFN was further selected for functional validation. Bioinformatics analysis showed that SFN was significantly upregulated in ICC tissues and closely associated with poor prognosis, advanced TNM stage, lymph node metastasis, perineural invasion, and non-response to HAIC combined with lenvatinib and PD-1 inhibitor therapy. SFN is particularly important in signal transduction, regulation of the cell cycle, and the cellular response to stress. In HCC, elevated SFN expression correlated with decreased overall survival and disease-free survival rates. SFN overexpression can activate the Wnt/β-catenin signaling pathway by decreasing the phosphorylation of β-catenin. This process facilitates the translocation of β-catenin into the nucleus, thereby increasing the expression of the c-myc oncogene [36]. Moreover, SFN overexpression mitigates sorafenib-induced effects on HCC cells. SFN enhances the activity of AKT kinase by interfering with the interaction between PHLPP2 and AKT, thus promoting HCC progression [37]. In ICC, high SFN expression was significantly associated with poor overall survival. Knocking down SFN in ICC cell lines reduced cell migration, invasion and anoikis resistance [38].

In vitro experiments demonstrated that SFN knockdown markedly inhibited the proliferation and colony formation ability of HuCCT1 and RBE cells, enhanced sensitivity to cisplatin and gemcitabine of HuCCT1 cell, and increased γ-H2AX expression, indicating increased number of DNA double-strand breaks. We also examined whether SFN knockdown affects the sensitivity of RBE cells to gemcitabine. The results showed that although SFN knockdown partially enhanced gemcitabine sensitivity, the IC₅₀ values remained relatively high, consistent with the inherent insensitivity of RBE cells to gemcitabine reported in previous studies [39]. These findings suggest that while SFN may play a role in modulating chemosensitivity, its effect under baseline conditions is limited, and the clinical significance of this observation warrants further investigation. Collectively, these results confirm that SFN acts as a key driver of ICC progression and chemotherapy resistance, representing a promising therapeutic target for ICC.

In this comprehensive study, we integrated machine learning algorithms with biological experiments to identify key prognostic genes associated with the DDR signature in ICC. By combining single-cell sequencing, bulk RNA sequencing and spatial transcriptomics, we identified key DDR genes that significantly impact the prognosis of ICC. These findings provide valuable insights into the molecular mechanisms of the DDR in ICC and highlight potential therapeutic targets for personalized treatment strategies.

However, our research has several limitations. The general conclusions were drawn from public datasets and were only preliminarily validated through in vitro experiments. Moreover, the model has not been applied or assessed within the context of actual clinical practice. In the future, we will maintain our dedication to closely monitoring these issues and diligently work toward addressing them.

Conclusion

Our study integrated machine learning and biological experiments to identify key prognostic genes associated with the DDR signature in ICC based on single-cell sequencing, bulk RNA sequencing and spatial transcriptomics data. The model can perfectly predict the prognosis, clinical features, immune microenvironment and drug sensitivity of ICC patients. While derived from public datasets and validated in vitro, these findings suggest that our DDR-related model could serve as a valuable tool for individualized risk assessment and treatment decision-making strategies for clinical doctors. Further real-world clinical trials are essential to confirm these results.

Abbreviations

AUC

Area under the curve

BER

Base excision repair

CCK-8

Cell counting kit-8

CIBERSORT

Immune infiltration analysis algorithm

DCA

Decision curve analysis

DDR

DNA damage response

DEGs

Differentially expressed genes

DSBs

Dna double-strand breaks

DCs

Dendritic cells

FBS

Fetal bovine serum

GO

Gene ontology

HAIC

Hepatic arterial infusion chemotherapy

HCC

Hepatocellular carcinoma

ICC

Intrahepatic cholangiocarcinoma

IC50

Half maximal inhibitory concentration

ICIs

Immune checkpoint inhibitors

KEGG

Kyoto encyclopedia of genes and genomes

KM

Kaplan − meier

LASSO

Least absolute shrinkage and selection operator

MMR

Mismatch repair

NER

Nucleotide excision repair

NHEJ

Nonhomologous end joining

OS

Overall survival

PCA

Principal component analysis

PD-1

Programmed death 1

PD-L1

Programmed death ligand 1

PFS

Progression-free survival

ROC

Receiver operating characteristic

RSF

Random survival fores

tscRNA-seq

Single-cell RNA sequencing

GSEA

Single-sample gene set enrichment analysis

SVMs

Support vector machines

TCGA

The cancer genome atlas

TME

Tumor microenvironment

t-SNE

T-distributed stochastic neighbor embedding

TNM

Tumor-node-metastasis

γ-H2AX

Phosphorylated histone H2AX

Author contributions

Kai Lu: Writing – original draft, Visualization, Data curation, Conceptualization. Hanqi Li: Data curation, Visualization and Validation. Yinying Wu: Writing – review & editing. Xuyuan Dong: Writing – review & editing. Danfeng Dong: Writing – review & editing. Yangwei Fan: Writing – review & editing. Enxiao Li: Investigation, Funding acquisition. Liankang Sun: Investigation, Funding acquisition. Yu Shi: Investigation, Funding acquisition. All authors reviewed the manuscript.

Funding

This study was supported by grants from the National Natural Science Foundation of China (82303663), the Natural Science Basic Research Program of Shaanxi (2022JQ-846), the Research Foundation of the First Affiliated Hospital of Xi’an Jiaotong University (2021QN-01,2022QN-15), and Fundamental Research Funds for the Central Universities (xzy012022103).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

All data utilized in this study were retrieved from publicly available online databases, as specified in the manuscript.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.