Integrative single-cell and machine-learning analysis identifies ac⁴C-related S100A13 as a causal risk gene in cholangiocarcinoma

Abstract

Background

N⁴-acetylcytidine (ac⁴C) is an emerging epitranscriptomic modification that regulates mRNA stability and translation, yet its biological and clinical relevance in cholangiocarcinoma (CCA) remains unclear.

Methods

We integrated single-cell RNA sequencing (scRNA-seq) and multi-cohort bulk transcriptomic data to systematically profile ac⁴C-related genes (acRGs) in CCA. Using UCell scoring and CellChat analysis, we assessed ac⁴C activity and intercellular communication within the tumor microenvironment (TME). A comprehensive machine learning framework combining 117 model algorithms was implemented to construct an ac⁴C-related gene signature (acRGS) for prognostic prediction. Immune contexture, causal inference, and in vivo validation were subsequently performed to elucidate functional relevance.

Results

Single-cell analysis revealed that malignant and myeloid populations exhibited the highest ac⁴C activity and denser ligand-receptor crosstalk. The derived 11-gene acRGS robustly stratified patients into prognostic groups across independent cohorts. High-risk tumors showed elevated checkpoint expression but markedly reduced immune and stromal infiltration, indicating an immune-exhausted yet poorly inflamed TME. Mendelian randomization identified S100A13 and ASPH as causal ac⁴C-linked risk genes for CCA. Functional experiments confirmed that S100A13 overexpression significantly enhanced tumor growth and proliferation in vivo.

Conclusions

Our integrative framework delineates the transcriptional and immunological consequences of ac⁴C modification in CCA and identifies S100A13 as a novel ac⁴C-associated oncogene. The acRGS provides a clinically relevant tool for prognostic assessment and mechanistic insight into RNA acetylation-driven tumor progression.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-026-15602-5.

Introduction

Cholangiocarcinoma (CCA) is a highly aggressive malignancy arising from the biliary epithelium and represents the second most common primary hepatic cancer [1–4]. Despite surgical advances, the majority of patients are diagnosed at advanced stages, with limited treatment options and a dismal 5-year survival rate of less than 20% [5–8]. Current systemic therapies, such as chemotherapy, molecular targeted therapy, and immune checkpoint blockade, offer only modest benefit due to profound intertumoral and intratumoral heterogeneity, leading to variable therapeutic responses and rapid development of resistance [9–13]. Understanding the molecular mechanisms underlying this heterogeneity is therefore crucial for improving prognostic stratification and identifying novel therapeutic targets in CCA.

Among emerging regulators of cancer biology, RNA modifications have gained increasing attention as epitranscriptomic determinants of gene expression [14–16]. More than 170 RNA chemical modifications have been identified, among which N⁴-acetylcytidine (ac⁴C) has recently emerged as a key player in maintaining mRNA stability and translation efficiency [17–19]. The modification is catalyzed by N-acetyltransferase 10 (NAT10), the only known ac⁴C “writer,” and is dynamically involved in diverse physiological and pathological processes [20–22]. Accumulating evidence suggests that aberrant ac⁴C modification contributes to tumorigenesis, progression, and therapy resistance across multiple cancer types [23–25]. However, its biological role and clinical relevance in cholangiocarcinoma remain largely unexplored.

In this study, we comprehensively characterized ac⁴C-related transcriptomic features in cholangiocarcinoma using integrated single-cell and bulk RNA sequencing (RNA-seq) analyses. We established an ac⁴C activity landscape at the single-cell level, constructed a robust machine-learning-based prognostic model, and explored its association with the tumor immune microenvironment. Furthermore, causal inference and experimental validation were performed to identify key ac⁴C-associated drivers, highlighting S100A13 as a potential oncogenic effector in CCA. Collectively, this work provides a systems-level understanding of ac⁴C modification in CCA and reveals its potential as both a biomarker and a therapeutic target.

Materials and methods

Data sources

Single-cell and bulk RNA-seq datasets were obtained from publicly available repositories. The single-cell RNA-seq dataset GSE138709 (cholangiocarcinoma, n = 8) was downloaded from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE138709) [26]. Four independent bulk cohorts were used for model development and validation: E-MTAB-6389 (https://www.ebi.ac.uk/biostudies/ArrayExpress/studies/E-MTAB-6389, n = 109), GSE244807 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE244807, n = 246) [27], GSE107943 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE107943, n = 57) [28], and TCGA-CHOL (The Cancer Genome Atlas, https://portal.gdc.cancer.gov/, n = 45). All data were processed using R software (v4.3.3) and associated Bioconductor and CRAN packages. Pan-cancer expression analysis was retrieved from the GEPIA database (http://gepia.cancer-pku.cn/) [29].

Single-cell RNA-seq data processing and integration

The single-cell transcriptomic dataset (GSE138709) was processed using the Seurat package. Cells with fewer than 200 detected genes, more than 5000 genes, or with > 10% mitochondrial or > 5% hemoglobin gene expression were excluded. Doublets were detected and removed using the DoubletFinder package. After quality control, a total of 27,628 cells were retained for downstream analysis. Cluster marker genes were identified using the FindAllMarkers function with parameters: logfc.threshold = 0.25, min.pct = 0.1, and p_val_adj < 0.05.

ac⁴C-related gene activity scoring and classification

A curated list of 2135 ac⁴C-related genes (acRGs) was collected from previously published ac⁴C modification databases and literature (Table. S1) [30]. Single-cell-level acRG activity was quantified using the UCell algorithm. Cells were then divided into acRG_high and acRG_low groups based on the median acRG score across all cells. UMAP visualization and violin plots were generated to illustrate acRG activity distribution across cell types and groups.

Cell–cell communication analysis

Intercellular communication networks were inferred using CellChat [31]. Normalized expression matrices from Seurat were used as input, and cell-type annotations were adopted as identities. Communication probability and interaction strength were estimated using the human ligand–receptor database provided by CellChat. Network statistics, including total interaction number and weight, were computed for both acRG_high and acRG_low groups.

Differential expression and correlation analyses

Differentially expressed genes (DEGs) between acRG_high and acRG_low groups were identified using Seurat’s FindMarkers function. Spearman correlations between each gene’s normalized expression and the acRG score were computed across all cells. Genes were ranked by correlation, and the top 150 correlated genes were selected. Intersecting these with DEGs yielded 145 acRG-associated genes used for downstream modeling.

Machine learning–based prognostic modeling

Bulk expression data from the four cohorts (E-MTAB-6389, GSE244807, GSE107943, TCGA-CHOL) were normalized to log₂(TPM + 1) before modeling. A comprehensive machine-learning framework was implemented using the Mime1 and randomForestSRC packages. The pipeline evaluated 117 combinations integrating various feature selection methods (e.g., univariate Cox, Stepwise Cox, LASSO) and survival learners (e.g., RSF, CoxBoost, plsRcox, Elastic Net) [32, 33]. Model performance was evaluated by C-index and time-dependent AUC across all cohorts. Survival analysis was conducted using Kaplan-Meier curves and the log-rank test, and group separation was visualized by PCA.

Immune checkpoint and tumor microenvironment analyses

Immune checkpoint gene (ICG) expression differences between high- and low-acRGS risk groups were evaluated using the Wilcoxon test. Spearman correlations were computed between acRGS model genes (and risk score) and ICGs. Tumor microenvironment composition was analyzed using multiple computational frameworks, including xCell, EPIC, and ESTIMATE [34–36]. All scores were compared between risk groups using Wilcoxon tests (adjusted p < 0.05).

Mendelian randomization analysis

To test causal relationships between acRGS genes and cholangiocarcinoma (CCA) risk, two-sample Mendelian randomization (MR) was conducted. Instrumental variables (single-nucleotide polymorphisms, SNPs) for each gene were extracted from the deCODE database (https://download.decode.is/form/folder/proteomics, n = 35559) using the criteria p < 5 × 10⁻⁸ and F-statistic > 10. Independent SNPs for each exposure were obtained by LD clumping with thresholds of R² < 0.001 and a physical distance window of 10,000 kb. The outcome summary statistics were retrieved from the GWAS Catalog, using the publicly available dataset GCST90018803 (https://www.ebi.ac.uk/gwas/studies/GCST90018803). Access to the deCODE proteomics GWAS data was obtained under a formal data-use agreement, whereas the GWAS Catalog dataset GCST90018803 (n = 476091) is fully open access and was used in compliance with its public data-use policy. Therefore, no individual-level data were accessed, and no permissions were required. We performed two-sample MR analyses using the inverse variance weighted (IVW), MR-Egger, weighted median, weighted mode, and simple mode methods. Heterogeneity was evaluated with Cochran’s Q test, and pleiotropy with Egger intercepts. Leave-one-out analysis was used for sensitivity testing. Genes with significant IVW p < 0.05 and consistent direction across estimators were considered causally associated with CCA. To verify the causal direction, we applied Steiger filtering to compare the variance explained (R²) by exposure versus the outcome. This Mendelian randomization study was designed and reported in accordance with the STROBE-MR guidelines for reporting Mendelian randomization analyses.

In vivo functional validation of S100A13

Human cholangiocarcinoma cell line HUCCT1 was transduced with lentivirus carrying S100A13 overexpression (LV-S100A13) or negative control (LV-NC) constructs. Stable cell lines were confirmed by qRT-PCR with GAPDH as internal control. For xenograft assays, BALB/c nude mice (4–6 weeks) were subcutaneously injected with 5 × 10⁶ HUCCT1 cells (LV-S100A13 or LV-NC) (n = 5). Tumor growth was monitored weekly for 4 weeks, and volume calculated as (length × width²)/2. Histological sections were stained with hematoxylin-eosin (H&E) and Ki-67 to assess morphology and proliferation. All animal procedures were approved by the Institutional Animal Care and Use Committee of our institution and were performed in accordance with national and international guidelines for the care and use of laboratory animals. For terminal experiments, animals were first deeply anaesthetised with isoflurane (4–5% for induction and 1.5-2.0% for maintenance in oxygen) until loss of pedal withdrawal and corneal reflexes was confirmed. While under deep anaesthesia, animals were euthanised by gradual-fill CO₂ inhalation in a dedicated chamber (CO₂ flow rate 20–30% of the chamber volume per minute) until respiratory arrest was observed, followed by a brief observation period to ensure irreversible death. This two-step protocol (general anaesthesia followed by CO₂ euthanasia) was chosen to minimise pain and distress and is consistent with current AVMA and institutional recommendations. No animals were euthanised without prior confirmation of deep anaesthesia, and all procedures were carried out by trained personnel.

Results

Analysis of cholangiocarcinoma at the single-cell level

The study workflow is outlined in Fig. 1. A total of 27,628 single cells from cholangiocarcinoma (GSE138709) were integrated and visualized using UMAP, revealing well-mixed distributions across samples and groups, clear clustering, and distinct lineage identities (Fig. 2A–D). Standard quality control removed low-quality cells and potential doublets before integration (Fig. S1). Canonical marker analysis confirmed accurate cell-type annotation across clusters (Fig. 2E). The relative proportions of major cell types varied among samples (Fig. 2F). Cell-cell communication analysis further identified dense interlineage networks, with visualization by both interaction number and interaction weight highlighting strong crosstalk between malignant and immune populations (Fig. 2G–H).

Fig. 1.

The flowchart

Fig. 2.

Integrated single-cell atlas, cell-type annotation, composition, and intercellular communication networks. A–D UMAP visualization of 27,628 single cells colored by sample, group, cluster, and annotated cell type, respectively. E Dot plot showing canonical marker gene expression across clusters. Dot size indicates the proportion of expressing cells, and color intensity represents average expression. F Stacked area plot showing relative proportions of each cell type across samples. G-H Cell–cell communication network visualized by interaction number and interaction weight

To investigate RNA modification–related heterogeneity at the single-cell level, we assessed the activity of 2135 ac⁴C-related genes (acRGs) using the UCell algorithm. Distinct cell types exhibited variable acRG activity, with malignant cells and macrophages showing relatively high scores (Fig. 3A–B). Marker gene patterns remained consistent across cell types (Fig. 3C), and UMAP visualization revealed spatially heterogeneous acRG activity within the tumor microenvironment (Fig. 3D).

Fig. 3.

ac⁴C-related gene (acRG) activity and its association with cellular communication in cholangiocarcinoma. A Bubble plot showing acRG activity evaluated by UCell using multiple scoring metrics. B Violin plot of acRG activity across cell types. C Dot plot showing marker gene expression by cell type and acRG group. D UMAP colored by acRG score distribution. E Bar plots showing the number and aggregate strength of inferred cell–cell interactions in acRG_low and acRG_high. F–G Heatmaps of interaction counts between source and target cell types in acRG_low and acRG_high, respectively. H–I Centrality scatter plots showing outgoing versus incoming interaction strength by cell type in acRG_low and acRG_high. J Correlation ranking of genes with the acRG score (Spearman). K Venn diagram showing the overlap between top-150 correlation-ranked genes and DEGs, yielding 145 candidates

Fig. 4.

Comparative cell–cell communication between acRG_high and acRG_low groups. A Global landscape of signaling pathways showing relative (left) and absolute (right) information flow inferred by CellChat in acRG_high (red) and acRG_low (blue) groups. B Top 25 pathways ranked by total information flow. C Differential communication (ΔFlow = high − low) highlighting the direction and magnitude of pathway changes. Bars represent aggregated ligand–receptor interaction probabilities across all cell types. Pathways with near-zero signaling in one group (e.g., LAMININ) appear visually absent on that side, reflecting minimal activity rather than missing data

The acRG_high group showed more inferred interactions and greater overall interaction strength (Fig. 3E), which was evident at the cell-type level by denser interaction matrices (Fig. 3F–G) and by higher incoming and outgoing centrality across multiple lineages (Fig. 3H–I). At the pathway level, acRG_high displayed stronger immune-modulatory signaling (e.g., MIF, MHC-I/II, COLLAGEN), with ΔFlow summarizing the net shifts between groups (Fig. 4A–C). Finally, genes were ranked by Spearman correlation with the acRG score and the top 150 candidates were intersected with DEGs between acRG_high and acRG_low (p_adj < 0.05; |log2FC| > log2(1.5); min.pct = 0.1), yielding 145 acRG-associated genes for downstream analyses (Fig. 3J–K).

Machine learning–based prognostic modeling

To identify ac⁴C-related prognostic features, we implemented a comprehensive machine-learning pipeline on the E-MTAB-6389 training cohort and validated performance in GSE244807, GSE107943, and TCGA-CHOL. In total, 117 combinations integrating feature selection methods and survival learners were evaluated. The StepCox[forward] + RSF model achieved the highest mean C-index across datasets and the best time-dependent AUC at 5 years (Figs. 1 and 5A–B- and 3-year AUCs in Fig. S2).

Fig. 5.

Machine-learning development and performance of the ac⁴C-related prognostic model. A Heatmap of C-index for 117 model combinations across the training cohort (E-MTAB-6389) and external validations (GSE244807, GSE107943, TCGA-CHOL). B Corresponding 5-year AUC across all models and cohorts. The StepCox[forward] + RSF model shows the best average performance across datasets. Color scales indicate metric values; right-side summary bars show cohort-level means

The resulting ac⁴C-related gene signature (acRGS) comprised 11 genes, including S100A13, PLAT, FSCN1, PLEC, ASPH, TNFRSF12A, AHNAK, NME1, APLP2, KTN1, and MTDH. Using the optimal model, a patient-level risk score separated individuals into high- and low-risk groups with significantly worse overall survival in the high-risk group in the training and GSE244807 cohorts, with consistent trends in TCGA and GSE107943 (Fig. 6A–D). Principal component analysis further showed clear transcriptomic separation between risk groups across cohorts (Fig. 6E–H). Collectively, these results indicate that the acRGS provides a stable and generalizable prognostic tool for cholangiocarcinoma. This cross-cohort benchmarking strategy is consistent with recent computational oncology studies that emphasize systematic model comparison and external validation for robust prognostic signature development [37–39].

Fig. 6.

Validation of the optimal model and risk-group separation. A–D Kaplan–Meier survival curves in E-MTAB-6389, GSE244807, TCGA-CHOL, and GSE107943, stratified by the acRGS risk score (median-based cut-off). E–H PCA plots with marginal densities showing transcriptomic separation between high- and low-acRGS groups across cohorts

Immune contexture associated with the acrgs risk signature

Patients in the GSE244807 cohort were divided into high and low acRGS risk groups based on the median risk score.

Immune checkpoint gene (ICG) analysis revealed that most checkpoint molecules were significantly up-regulated in the high-risk group (Wilcoxon test; multiple genes p < 0.05), suggesting an immune-activated yet potentially exhausted tumor state (Fig. 7A).

Fig. 7.

Immune contexture associated with acRGS risk. A Expression of immune-checkpoint genes between high and low acRGS risk groups. B Spearman correlations between model genes (and risk score) and checkpoint genes. C xCell-inferred cell-type abundance showing reduced dendritic cells, M1 macrophages, and monocytes in the high-risk group. D EPIC-estimated CD4⁺ T-cell fractions decreased in the high-risk group. E ESTIMATE results showing lower StromalScore, ImmuneScore, and ESTIMATEScore, but higher TumorPurity in the high-risk group

Spearman correlation analysis demonstrated coherent associations between acRGS model genes (plus risk score) and checkpoint expression, indicating both positive and negative coupling patterns (Fig. 7B).

Deconvolution of the tumor microenvironment using xCell revealed that high-risk tumors contained fewer dendritic cells (DCs), M1 macrophages, and monocytes (Fig. 7C). Similarly, EPIC analysis showed a lower CD4⁺ T-cell fraction in the high-risk group (Fig. 7D). Consistent with these observations, the ESTIMATE algorithm indicated lower StromalScore, ImmuneScore, and ESTIMATEScore, and higher TumorPurity in the high-risk group (Fig. 7E). Together, these results suggest that high acRGS risk corresponds to a less inflamed and less stromal tumor microenvironment, characterized by reduced immune infiltration despite elevated checkpoint expression. The convergence of multiple deconvolution frameworks on reduced immune/stromal components alongside elevated checkpoint expression further supports a decoupling between checkpoint activation and effective immune infiltration, a phenomenon also noted in immune profiling studies of hepatobiliary inflammatory settings [40].

Causal inference and experimental validation of S100A13 as an ac⁴C-related risk gene in cholangiocarcinoma

To test causality, we performed two-sample Mendelian randomization (MR) for each of the 11 risk genes defined by the acRGS machine-learning signature. Across standard MR estimators (IVW, MR-Egger, weighted median/mode), ASPH and S100A13 showed significant positive associations with cholangiocarcinoma risk under the IVW model (ASPH: p = 0.02; S100A13: p = 0.03), while the other risk genes did not reach significance (Fig. 8A). Leave-one-out analyses indicated that no single instrument drove the signals (Fig. 8B), and method-consistent positive slopes supported robustness of the effects (Fig. 8C) (Table. S2-4). These results suggest that ASPH and S100A13 may act as causal mediators linking the acRGS program to CCA susceptibility. Consistent with the Mendelian randomization findings, both ASPH and S100A13 were significantly upregulated in cholangiocarcinoma compared with normal tissues across pan-cancer datasets (Fig. 8D–E). Such transcriptional upregulation may reflect upstream regulatory circuits involving non-coding RNAs, as suggested by prior studies linking lincRNA-miRNA networks to metastatic and oncogenic processes [41].

Fig. 8.

Functional validation of S100A13 in cholangiocarcinoma. A–C Mendelian randomization (MR) analyses of acRGS genes identified S100A13 and ASPH as causally associated with cholangiocarcinoma (CCA) risk. D–E Pan-cancer expression profiles from the GEPIA database showing consistent upregulation of ASPH (D) and S100A13 (E) in CCA relative to normal tissues. F–J Functional validation of S100A13 in vivo

Among the machine learning-derived acRGS genes, both S100A13 and ASPH showed significant causal associations with cholangiocarcinoma risk in Mendelian randomization analysis. While ASPH has been sporadically reported in biliary malignancies, S100A13 remains largely uncharacterized in cholangiocarcinoma. To experimentally evaluate its potential oncogenic role, we established S100A13-overexpressing HUCCT1 cells (LV-S100A13), which displayed markedly elevated mRNA levels compared with control cells (LV-NC) (Fig. 8F) (Table. S5). In vivo xenograft assays demonstrated that S100A13 overexpression significantly promoted tumor growth, as evidenced by increased tumor size and volume in nude mice (Fig. 8G–H). Histological analysis revealed more aggressive morphology in the LV-S100A13 group, accompanied by enhanced Ki-67 staining, indicating higher proliferative activity (Fig. 8I–J). Collectively, these findings provide preliminary experimental evidence that S100A13 acts as a tumor-promoting factor in cholangiocarcinoma, consistent with its genetic and transcriptomic associations identified in preceding analyses.

Discussion

RNA acetylation, particularly N⁴-acetylcytidine (ac⁴C) catalyzed by N-acetyltransferase 10 (NAT10), has emerged as a pivotal layer of post-transcriptional regulation in cancer biology. This modification enhances mRNA stability and translation efficiency, thereby modulating fundamental cellular processes such as proliferation, differentiation, metabolism, and stress responses [42–44]. Dysregulated ac4C modification has been implicated in multiple malignancies, such as cervical, gastric, and colorectal cancers, where it stabilizes oncogenic transcripts and promotes malignant progression. Despite accumulating evidence linking ac⁴C to tumorigenesis, its role in cholangiocarcinoma (CCA) has remained largely unexplored.

In this study, we systematically delineated the role of ac⁴C-related genes (acRGs) in CCA through integrative analysis of single-cell and bulk transcriptomes. Single-cell RNA-seq profiling revealed that ac⁴C activity varied substantially across cellular compartments, with malignant epithelial cells and macrophages exhibiting the highest acRG scores. The acRG_high phenotype was characterized by denser intercellular communication networks and enhanced immune-modulatory signaling, indicating that ac⁴C modification may contribute to tumor-immune crosstalk and microenvironmental remodeling [45, 46]. Notably, the dominance of immunomodulatory ligand-receptor programs (e.g., MIF/MHC/COLLAGEN) in acRG_high cells echoes the broader concept that remodeling of intercellular crosstalk can shape immune contexture and therapeutic responsiveness in complex inflammatory and tumor settings [47]. From a systems-level perspective, our integrative workflow is conceptually aligned with recent multi-omics integration studies [48]. Here, we extend this systems-level paradigm to epitranscriptomic regulation and tumor microenvironment remodeling in cholangiocarcinoma.

By integrating these insights with multi-cohort bulk transcriptomic data, we established a robust machine learning–derived ac⁴C-related gene signature (acRGS) comprising 11 prognostic genes. The acRGS effectively stratified patients into high- and low-risk groups with significantly distinct survival outcomes, demonstrating strong and reproducible predictive power across independent datasets. Consistent with prior biomarker optimization studies, the use of a multi-gene signature enables incremental performance gains over single-gene predictors by capturing coordinated network-level effects rather than isolated gene signals [49]. Importantly, immune deconvolution revealed a striking paradox: the high-acRGS group exhibited elevated expression of immune checkpoint molecules (e.g., PDCD1, CTLA4, LAG3) despite reduced immune and stromal infiltration. A similar checkpoint-high/immune-low pattern has been reported in intrahepatic cholangiocarcinoma, where stromal exclusion mechanisms were implicated in T-cell dysfunction and immune escape [50]. This pattern reflects an immune-exhausted or immune-desert phenotype, where checkpoint upregulation signifies chronic antigenic stimulation and T-cell dysfunction rather than effective immune activation [51]. Such tumors often maintain an immunosuppressive milieu characterized by stromal exclusion and functional T-cell paralysis, which facilitate immune evasion and therapeutic resistance. These observations underscore the multifaceted influence of ac⁴C on both tumor-intrinsic aggressiveness and the surrounding immune contexture in CCA. Conceptually, this myeloid-enriched and immunosuppressive contexture is consistent with immune remodeling programs described in other malignancies [52], suggesting that ac4C-associated regulation may contribute to macrophage polarization and downstream TME immunosuppression in CCA. Methodologically, the refinement of acRG candidates through correlation ranking combined with differential expression analysis reflects a pathway-oriented and reproducible feature engineering strategy, which has also been applied in recent studies of cholangiocarcinoma biology and therapeutic resistance [53, 54].

To probe potential causality, we conducted two-sample Mendelian randomization (MR) analysis across the acRGS genes [55]. This approach identified S100A13 and ASPH as putative ac⁴C-driven causal mediators of CCA susceptibility, with consistent directionality across MR estimators. While ASPH has been sporadically linked to biliary malignancies, S100A13 remains largely uncharacterized in CCA. This genetic prioritization strategy is consistent with prior studies that used causal inference to nominate candidate regulatory networks and generate mechanistic hypotheses in oncology [56]. Our in vivo experiments demonstrated that overexpression of S100A13 markedly accelerated tumor growth and proliferation in xenograft models, supported by increased tumor volume, higher mitotic activity, and enhanced Ki-67 staining. Together, these findings provide the first experimental evidence implicating S100A13 as a tumor-promoting factor in CCA, bridging genetic, transcriptomic, and functional evidence under the ac⁴C regulatory framework.

Collectively, this integrative analysis establishes a comprehensive connection between epitranscriptomic regulation, immune contexture, and tumor progression in cholangiocarcinoma. The acRGS not only provides mechanistic insights into ac⁴C-associated transcriptional programs but also offers a clinically actionable prognostic model that may inform immunotherapy stratification.

Looking forward, the integrative framework presented in this study provides a foundation for future methodological and translational advances in cholangiocarcinoma research. Recent developments in machine learning, artificial intelligence, and mathematical modeling have increasingly enabled the integration of high-dimensional multi-omics data to capture tumor heterogeneity, predict clinical outcomes, and optimize therapeutic decision-making [57–59]. In this context, ac4C-related transcriptional programs and the acRGS defined here could be incorporated into more advanced predictive models, including deep learning–based survival frameworks, network-informed risk stratification, or dynamical models of tumor–immune interactions. Such approaches may facilitate personalized prognostic assessment, improve immunotherapy response prediction, and support rational combination strategies targeting both epitranscriptomic regulation and immune checkpoints. Together, these perspectives highlight the potential of combining epitranscriptomic profiling with emerging computational paradigms to further advance precision oncology in cholangiocarcinoma.

However, several limitations should be acknowledged. First, ac⁴C activity was inferred from ac⁴C-related gene sets rather than directly measured by ac⁴C-mapping assays (e.g., acRIP-seq, ac⁴C-seq). Thus, the acRG score serves as an indirect proxy that may not fully capture site-specific acetylation dynamics. Second, the study did not experimentally perturb or quantify NAT10 or other potential “writers,” “readers,” or “erasers,” leaving enzyme-level regulatory mechanisms unresolved. Third, bulk-cohort analyses relied on computational deconvolution for tumor microenvironment estimation, which may introduce model-dependent bias. Fourth, although MR inference was robust across estimators, it remains susceptible to weak-instrument and horizontal pleiotropy effects. Finally, functional validation focused primarily on S100A13; comprehensive gain- and loss-of-function studies across multiple cell models, as well as ASPH perturbation and direct ac⁴C enzyme inhibition experiments, are warranted to confirm causality. Future efforts integrating spatial transcriptomics, epitranscriptome-wide mapping, and prospective clinical cohorts will be essential to refine the mechanistic and translational significance of ac⁴C in CCA.

Conclusions

Through integrated single-cell, multi-cohort, and machine learning analyses, this study provides a comprehensive overview of ac⁴C modification in cholangiocarcinoma. We established an ac⁴C-related prognostic gene signature (acRGS) with strong predictive performance and uncovered its close link to the tumor immune microenvironment. Causal and experimental validation identified S100A13 as a novel ac⁴C-associated oncogene that promotes CCA progression. Together, these findings reveal ac⁴C modification as a key epitranscriptomic mechanism shaping tumor behavior and immune landscape in CCA, offering new opportunities for biomarker development and therapeutic intervention.

Abbreviations

ac⁴C

N⁴—acetylcytidine

CCA

Cholangiocarcinoma

acRG

ac⁴C—related genes

acRGS

acRG—related prognostic signature

scRNA

seq—Single—cell RNA sequencing

TME

Tumor microenvironment

UMAP

Uniform Manifold Approximation and Projection

DEG

Differentially expressed gene

UCell

Single—sample gene—set scoring method for scRNA—seq

CellChat

Cell—cell communication inference framework

ML

Machine learning

PCA

Principal component analysis

KM

Kaplan—Meier

OS

Overall survival

C

index—Concordance index

AUC

Area under the ROC curve

RSF

Random survival forest

CoxBoost

Likelihood—based boosting for Cox models

plsRcox

Partial least squares regression Cox model

LASSO

Least absolute shrinkage and selection operator

MR

Mendelian randomization

IVW

Inverse—variance weighted (MR estimator)

SNP

Single—nucleotide polymorphism

eQTL

Expression quantitative trait locus

GTEx

Genotype—Tissue Expression project

TCGA

The Cancer Genome Atlas

GEO

Gene Expression Omnibus

GEPIA

Gene Expression Profiling Interactive Analysis

EPIC

Estimating the Proportions of Immune and Cancer cells

xCell

Gene—signature—based cell type enrichment analysis

ESTIMATE

Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data

TPM

Transcripts per million

HR

Hazard ratio

CI

Confidence interval

H&E

Hematoxylin and eosin (staining)

IHC

Immunohistochemistry

NAT10

N—acetyltransferase 10

DC

Dendritic cell

Authors’ contributions

YZ and YL contributed equally to this work and are co-first authors. YZ conceived the study, performed single-cell RNA-seq data analysis, implemented the machine learning framework, conducted Mendelian randomization analysis, and drafted the manuscript. YL curated ac⁴C-related gene sets, carried out bulk transcriptomic validation, performed immune microenvironment deconvolution, and assisted in statistical analysis and figure preparation. ZW contributed to data preprocessing, functional enrichment analysis, and literature review. FW supervised the entire project, designed the experimental validation strategy, interpreted the biological and clinical implications, and critically revised the manuscript. All authors read and approved the final version of the manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). All animal experiments were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of The First Affiliated Hospital of Zhejiang University School of Medicine (Approval No. 2025 − 227). All methods were carried out in accordance with relevant guidelines and regulations. All methods are reported in accordance with ARRIVE guidelines (https://arriveguidelines.org).

Competing interests

The authors declare no competing interests.