Editing regulatory network in non-small cell lung cancer: implications for immunity and metastasis Authors: Jiajin Zhang, Liyu Huang, Sijia Wu Published: 2026 ABSTRACT BackgroundRNA editing plays critical roles in cancer development, and multiple editing events collectively exert greater effects than individual events. However, the coordinated impact of RNA editing remains poorly understood.MethodsA comprehensive framework was proposed to investigate interactions among RNA editing events, capturing both the global editing landscape and site-specific functional effects. The resulting RNA editing modules were further explored to reveal gene interactions and regulatory mechanisms from the perspective of RNA editing.ResultsUsing non-small cell lung cancer as an example, we identified more than 12 modules involved in immune regulation and metastasis-related processes, demonstrating their clinical relevance. Mechanistically, RNA editing influenced miRNA-mediated regulation of EIF2AK2 and PRKAR2A, transcription factor activity of AHR, and RNA-binding regulation of DDX3X, highlighting its role in shaping gene regulatory networks.ConclusionsCollectively, this study provides a systematic, module-based framework for understanding coordinated RNA editing, offering new insights into RNA editing–mediated gene interactions and revealing a complementary approach to study molecular mechanisms of diseases.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40246-026-00999-8. Introduction RNA editing, a critical post-transcriptional modification, has been increasingly recognized for its roles in cancer biology [1]. In particular, it has been implicated in non-small cell lung cancer (NSCLC) [2–4], one of the most commonly diagnosed malignancies and the leading cause of cancer-related mortality worldwide [5]. Specifically, RNA editing within double-stranded RNA (dsRNA) structures plays a pivotal role in preventing aberrant activation of innate immune responses [6]. Moreover, a single editing event at the RNA level can modify metastasis-promoting or -suppressing factors, thereby contributing to tumor invasion and progression [4, 7–9]. Consequently, elucidating the functional consequences and regulatory pathways of RNA editing is essential for uncovering novel mechanisms underlying cancer development and progression. To date, numerous studies have investigated the functional roles of RNA editing through both computational analyses and biological experiments [10, 11]. However, these efforts have primarily focused on the effects of individual RNA editing events. For instance, an RNA editing site in the 3’-UTR of EIF2AK2 has been shown to modulate its gene expression, as demonstrated by transfection and luciferase reporter assays [12]. According to the REDIportal database [13], this gene harbors dozens of editing sites which may collectively reshape post-transcriptional regulation [14] to modify gene–gene interaction networks. Given this complexity, investigating an RNA editing interaction network may offer a more comprehensive understanding of RNA editing roles in tumorigenesis and cancer progression. Thus, to elucidate the functional effects of widespread RNA editing, a comprehensive RNA editing network was constructed. This network was subsequently organized into distinct editing modules. The biological functions, regulatory interactions, and potential mechanisms associated with these modules in the context of NSCLC were then systematically investigated. The following sections describe this process in detail. Results Strong effects of multiple RNA editing events Prior to constructing and annotating the whole RNA editing network, a preliminary experiment was conducted to demonstrate the pronounced effects of multiple RNA editing events compared to individual ones. Considering the previously validated effect of a single editing site on EIF2AK2 expression and the presence of multiple editing sites within this gene, EIF2AK2 was used as an example to highlight the significance and necessity of studying RNA editing within an interaction network. In total, 54 RNA editing sites occurring in more than 20% of samples were identified in the 3’-UTR of EIF2AK2 from NSCLC patients in TCGA [10]. Individually, each editing event demonstrated predictive power for EIF2AK2 expression (Fig. 1a, Table S1), including the previously validated site (Figure S1). Collectively, these events altered 90 miRNA interactions, as assessed by the TargetScan [15] and miRanda [16] method, which may interfere in the regulation of EIF2AK2 expression as well as other genes. To further evaluate their combined effects, all events were incorporated into a multivariate linear regression model using ordinary least squares with 100 iterations of 10-fold cross-validation (Fig. 1b). Within this integrative model, individual editing events interacted with one another and exerted either positive or negative effects on EIF2AK2 expression (Fig. 1c). Compared with the best single-event predictor identified during cross-validation, the combined use of multiple RNA editing events resulted in a lower mean squared error (Fig. 1d). Consequently, the predicted EIF2AK2 expression based on multiple RNA editing events closely matched the observed expression (Fig. 1e). Fig. 1Potential combined effects of multiple RNA editing events on EIF2AK2 expression. a Associations between top-ranked RNA editing events and EIF2AK2 expression assessed using linear regression models. b Overview of the process to construct a multivariate RNA editing model for EIF2AK2 expression prediction. c Top ten RNA editing events significantly associated with EIF2AK2 expression ranked by absolute regression coefficients in multivariate models. d Performance comparison between optimal univariate and multivariate models, evaluated using a t-test. e Correlation between EIF2AK2 expression predicted by the multivariate model and the observed expression values Potential combined effects of multiple RNA editing events on EIF2AK2 expression. a Associations between top-ranked RNA editing events and EIF2AK2 expression assessed using linear regression models. b Overview of the process to construct a multivariate RNA editing model for EIF2AK2 expression prediction. c Top ten RNA editing events significantly associated with EIF2AK2 expression ranked by absolute regression coefficients in multivariate models. d Performance comparison between optimal univariate and multivariate models, evaluated using a t-test. e Correlation between EIF2AK2 expression predicted by the multivariate model and the observed expression values Additionally, RNA editing events of transcription factors may also modulate the activity of their downstream targets [17]. The aryl hydrocarbon receptor (AHR) is a transcription factor that has been implicated in unfavorable prognosis and therapeutic sensitivity in lung adenocarcinoma [18]. Our data further demonstrate that multiple RNA editing events in AHR collectively exert a stronger regulatory effect on the expression of its downstream tumor suppressor target, MFSD2A [19] (Figure S2). Furthermore, the expression of the key editing enzyme ADAR, which reflects the global RNA editing activity within a sample, also plays an important role in cancer biology. Specifically, ADAR expression was markedly elevated in tumor tissues compared with normal controls (Figure S3A), was significantly associated with immune cell infiltration (Figure S3B), and was higher in patients with lymph-node–positive disease than in those without lymph node involvement (Figure S3C). These observations suggest that RNA editing activity contributes to immune-related processes and metastatic progression in cancer. Additionally, the distribution of RNA editing sites within regulatory regions associated with key factors, including RBPs and miRNA target sites, was systematically analysed, as shown in Figure S4. A total of 932 events overlapped RBP regulatory regions derived from CLIP–seq data in the RBPsuite database [20], and 4,826 events were located within predicted miRNA regulatory regions annotated by TargetScan v8.0 [15]. Moreover, 533 RNA editing events were identified within genes encoding these regulatory factors, as well as transcription factors from the ReMap database [21]. These widespread enrichments indicate that RNA editing frequently occurs within shared regulatory contexts, with the potential to simultaneously influence multiple genes and regulatory pathways. Together, these findings indicate that RNA editing functions in a coordinated manner to shape gene regulation and critical cancer-associated pathways. In the following sections, RNA editing is examined at the network or module level to reveal its potential biological roles. RNA Editing regulatory network reveals high relevance to clinical phenotypes Given the importance of interactions among multiple RNA editing events, an RNA editing regulatory network was constructed using editing events from NSCLC tumor samples in TCGA (Fig. 2a). To ensure robustness, only nodes with high PageRank values and edges with strong Pearson correlations were retained. As a result, the network comprised 1,657 RNA editing events and 919,786 potential interactions, representing coordinated relationships among editing sites potentially co-regulated through shared transcription factor occupancy, RBP interactions, and miRNA-mediated regulatory mechanisms (Figure S4). Importantly, after adjusting for host gene expression using multiple regression models, 99.96% (918,971/919,331) of the interactions with available expression data remained statistically significant. It indicated that these associations were not solely driven by gene co-expression effects. These interactions involved 361 genes, of which 81% were dysregulated in tumor (Fig. 2b). This editing network captures RNA editing–mediated gene interactions, some of which will be investigated in detail in subsequent analyses. Fig. 2Key RNA editing modules in lung cancer. a Workflow illustrating the construction of RNA editing modules. b Fraction of edited genes showing differential expression between tumor and healthy controls, assessed using a t-test. c Heatmap of RNA editing modules associated with clinical outcomes, including overall survival evaluated using Cox regression and Kaplan–Meier analysis and pathological N stage assessed using a t-test. Pathologic N stage refers to lymph node metastasis status determined by pathological examination Key RNA editing modules in lung cancer. a Workflow illustrating the construction of RNA editing modules. b Fraction of edited genes showing differential expression between tumor and healthy controls, assessed using a t-test. c Heatmap of RNA editing modules associated with clinical outcomes, including overall survival evaluated using Cox regression and Kaplan–Meier analysis and pathological N stage assessed using a t-test. Pathologic N stage refers to lymph node metastasis status determined by pathological examination The whole RNA editing network was subsequently clustered into 22 modules using a random walk algorithm (Fig. 2a) Across these modules, only a limited proportion of genes were shared, suggesting that they are driven by largely independent biological mechanisms (Figure S5). Notably, the principal components of most modules exhibited strong associations with clinical phenotypes, including survival outcomes and disease stages (Fig. 2c). These modules were also found to participate in immune responses, replication, metabolism, and so on (Figure S6). It highlighted the involvement of RNA editing modules in the key cancer-related mechanisms [22–24]. Collectively, these RNA editing modules may serve as informative indicators of clinical outcomes and warrant further investigation to elucidate their roles in tumorigenesis. RNA editing regulatory network is associated with immune infiltration The top enriched immune-related pathways (Figure S6) and the strong associations between RNA editing and immune features [6] prompted us to investigate the RNA editing–immune regulatory network. Our analyses identified 12 immune-related modules (Figure S7). Notably, module 14 (Fig. 3a) comprised 92 editing events in 44 genes. They were significantly enriched in pathways associated with antiviral defense, wound response, immune effector functions, and innate immune responses (Fig. 3b). Fig. 3Biological functions and clinical relevance of Module 14. a Network diagram illustrating interactions among genes harboring RNA editing events within Module 14. Nodes represent genes, and edges indicate correlations between their editing events, with edge thickness proportional to correlation strength. b Gene Ontology Biological Process (GO–BP) enrichment analysis of Module 14, performed using over-representation analysis in NetworkAnalyst. c Associations between immune cell infiltration levels, estimated using ImmuCellAI, and the first principal component (PC1) of editing events in Module 14, assessed using Pearson correlation. Tex: exhausted T cells, iTreg: induced regulatory T cells, nTreg: natural regulatory T cells. d Comparison of PC1 values between patients without lymph node metastasis (N = 0) and those with lymph node metastasis (N > 0), evaluated using a t-test. e Kaplan–Meier survival analysis stratified by the PC1 of Module 14 Biological functions and clinical relevance of Module 14. a Network diagram illustrating interactions among genes harboring RNA editing events within Module 14. Nodes represent genes, and edges indicate correlations between their editing events, with edge thickness proportional to correlation strength. b Gene Ontology Biological Process (GO–BP) enrichment analysis of Module 14, performed using over-representation analysis in NetworkAnalyst. c Associations between immune cell infiltration levels, estimated using ImmuCellAI, and the first principal component (PC1) of editing events in Module 14, assessed using Pearson correlation. Tex: exhausted T cells, iTreg: induced regulatory T cells, nTreg: natural regulatory T cells. d Comparison of PC1 values between patients without lymph node metastasis (N = 0) and those with lymph node metastasis (N > 0), evaluated using a t-test. e Kaplan–Meier survival analysis stratified by the PC1 of Module 14 The immune infiltration analysis further confirmed the strong associations between the module and immune activity. Specifically, the first principal component of editing events in module 14 was negatively correlated with the infiltration of exhausted and regulatory T cells, as assessed using the ImmuCellAI tool [25] (Fig. 3c). These editing events likely interacted with one other to enhance the functions of immune-related genes and immune cells. This hypothesis was supported by the reduced metastasis (Fig. 3d) and improving survival outcomes in patients with high editing activity in module 14 (Fig. 3e). In contrast, editing events in several other modules may play opposing roles in immune activity and clinical phenotypes (Fig. 2c and Figure S7). For example, high editing activity in another immune-related module 32 (Figure S8a-b) was associated with increased infiltration of exhausted and regulatory T cells (Figure S8c), which in turn correlated with enhanced metastasis (Figure S8d) and poorer survival outcomes (Figure S8e). The interaction between the editing events in these two modules further strengthened the association between RNA editing and immune activity (Figure S9). A closer examination of these two modules revealed 32 immune-related genes from the ImmPort [26] and InnateDB [27] databases. Representative examples included the well-known EIF2AK2 [28]. Its potential regulatory mechanisms within the immune network were explored in the following section. Potential regulatory mechanisms driving immune-related RNA editing network The immune-related RNA editing network highlighted hub genes and key gene interactions mediated by RNA editing. For instance, the editing events of the well-known immune-related gene EIF2AK2 were strongly associated with those of PRKAR2A (Fig. 4a), suggesting a potential interaction between these two genes. Increased RNA editing of EIF2AK2 was associated with lower RNA editing PC scores, which in turn correlated with elevated EIF2AK2 expression (Fig. 4b). Mechanistically, this effect may be partially mediated through the disruption of strong canonical 8mer miRNA-binding motifs caused by RNA editing (Fig. 4b–c). Similarly, RNA editing of PRKAR2A was associated with increased gene expression, potentially through attenuation of miRNA-mediated post-transcriptional repression (Fig. 4b–c). Although no previous studies have reported a direct molecular interaction between these genes, their strong association and immune-related functions [28–30] indicate potential cooperative roles in immune regulation. Fig. 4Gene interactions revealed by RNA editing module analysis. a Correlation between RNA editing events in EIF2AK2 and PRKAR2A. Numbers in parentheses indicate the loading coefficients of individual editing events contributing to the corresponding RNA editing PC. b Associations between the principal components (PCs) of RNA editing events and the expression levels of their corresponding host genes, as well as correlations between the PCs of the two genes and between their expression levels, assessed using Pearson correlation. c RNA editing–mediated modulation of miRNA regulation for both genes. Putative miRNA-binding sites affected by RNA editing were predicted using TargetScan and miRanda, with emphasis on canonical 8mer seed matches. d Potential regulation of RNA editing events in EIF2AK2 and PRKAR2A by DDX3X. e Associations between DDX3X expression and the PCs of RNA editing events, assessed using Pearson correlation. Mediation analysis indicates a significant effect of DDX3X on both genes through RNA editing. f Differences in RNA editing levels between high and low DDX3X expression groups, evaluated using a t-test. g Expression levels of DDX3X, EIF2AK2, and PRKAR2A across the interferon-γ dominant immune subtype (C2) and other immune subtypes compared using a t-test Gene interactions revealed by RNA editing module analysis. a Correlation between RNA editing events in EIF2AK2 and PRKAR2A. Numbers in parentheses indicate the loading coefficients of individual editing events contributing to the corresponding RNA editing PC. b Associations between the principal components (PCs) of RNA editing events and the expression levels of their corresponding host genes, as well as correlations between the PCs of the two genes and between their expression levels, assessed using Pearson correlation. c RNA editing–mediated modulation of miRNA regulation for both genes. Putative miRNA-binding sites affected by RNA editing were predicted using TargetScan and miRanda, with emphasis on canonical 8mer seed matches. d Potential regulation of RNA editing events in EIF2AK2 and PRKAR2A by DDX3X. e Associations between DDX3X expression and the PCs of RNA editing events, assessed using Pearson correlation. Mediation analysis indicates a significant effect of DDX3X on both genes through RNA editing. f Differences in RNA editing levels between high and low DDX3X expression groups, evaluated using a t-test. g Expression levels of DDX3X, EIF2AK2, and PRKAR2A across the interferon-γ dominant immune subtype (C2) and other immune subtypes compared using a t-test To further investigate the mechanisms underlying the coordinated RNA editing of these two genes, candidate RBPs were identified through an integrative strategy. Specifically, RBPs were required to (i) have binding sites located within ± 500 bp of the RNA editing sites in PRKAR2A and EIF2AK2 (Fig. 4d), based on starBase [31]; (ii) show significant associations with both the RNA editing patterns of these two genes in Module 14 and their expression levels; and (iii) possess known functions relevance to RNA editing. This approach led to the identification of DDX3X as a putative common regulator of PRKAR2A and EIF2AK2. DDX3X (DEAD-box helicase 3, X-linked) is known to unwind complex RNA secondary structures, thereby exposing double-stranded RNA regions to ADAR1 and facilitating ADAR1-mediated RNA editing [32]. It may therefore modulate the editing of both EIF2AK2 and PRKAR2A (Fig. 4e-f), ultimately influencing their expression through RNA editing–mediated mechanisms. Notably, this regulatory hypothesis is supported by the observed differential expression of these three genes in the interferon-γ–dominant immune subtype (C2; Fig. 4g), consistent with their previously reported roles in interferon-related pathways [29, 32, 33]. Together, these findings provide an illustrative example of a potential mechanism linking coordinated RNA editing to immune-related biological functions. Dysregulation of RNA editing network in lymph node metastasis RNA editing networks were enriched in immune-related functions [34], as described above, as well as pathways involved in RNA and DNA metabolism [35], stimulus responses [36], and cellular replication [37] (Figure S6). These pathways were closely linked to metastatic processes. Consistently, RNA editing enzymes were differentially expressed (Figure S3c), and specific modules exhibited aberrant editing patterns in patients with lymph node metastasis (Fig. 3d and Figure S8d). To further assess the relevance of RNA editing to metastasis, RNA editing data were integrated with clinical information and gene expression profiles for lymph node metastasis prediction (Fig. 5a). Incorporating RNA editing alongside clinical data significantly improved the prediction performance (Fig. 5b), with an effect comparable to that achieved by adding gene expression. Furthermore, combining all three data types yielded a modest additional improvement over models based on clinical and gene expression data alone (Fig. 5b–c). Together, these results indicate that RNA editing provides substantial and complementary information relevant to metastatic progression. Fig. 5Contribution of RNA editing modules to the prediction of lymph node metastasis in non-small cell lung cancer. a Workflow illustrating the prediction of lymph node metastasis with and without RNA editing features. b Radar chart displaying the average area under the curve (AUC) values of the prediction model across 20 dataset combinations using multiple methods. c Performance comparison of the NeuralNetFastAI model with and without RNA editing features, evaluated using DeLong’s test. Vertical black lines indicate the mean differences in AUC values between models with and without RNA editing features across 20 repeated runs Contribution of RNA editing modules to the prediction of lymph node metastasis in non-small cell lung cancer. a Workflow illustrating the prediction of lymph node metastasis with and without RNA editing features. b Radar chart displaying the average area under the curve (AUC) values of the prediction model across 20 dataset combinations using multiple methods. c Performance comparison of the NeuralNetFastAI model with and without RNA editing features, evaluated using DeLong’s test. Vertical black lines indicate the mean differences in AUC values between models with and without RNA editing features across 20 repeated runs Further analysis identified 13 metastasis-associated editing modules (Figure S10). Notably, in metastasis-associated module 25 (Fig. 6a), patients with lymph node metastasis exhibited significantly lower RNA editing frequencies (Fig. 6b), which were associated with poorer survival outcomes (Fig. 6c). The hub genes of this module included AHR, a transcription factor known to mediate cancer stem cell properties and to promote cell proliferation, invasion, and migration [38]. Its editing was also aberrant in patients with positive lymph node (Fig. 6d). According to the TRRUST [39] and BCSCdb [40] databases, AHR regulates 19 downstream stem cell–related genes (Fig. 6e), including MFSD2A (Fig. 6f). Reduced editing of AHR was associated with decreased MFSD2A expression (Fig. 6g), which in turn facilitated the maintenance of stem cell properties (Fig. 6h) [41] and ultimately contributed to lymph node metastasis (Fig. 6i). Fig. 6RNA editing module 25 contributes to lymph node metastasis. a Network diagram illustrating interactions among genes harboring RNA editing events within Module 25. b Comparison of Module 25 PC1 between patients with lymph node metastasis (N > 0) and those without (N = 0), assessed using a t-test. c Kaplan-Meier survival analysis stratified by the PC1 of Module 25. d Comparison of RNA editing levels of the hub gene AHR between patients with lymph node metastasis (N > 0) and those without (N = 0), evaluated using a t-test. e The transcription factor AHR has 22 manually curated target genes in TRRUST v2, of which 86.36% are annotated as stemness-related genes in BCSCdb. f One AHR-regulated stemness-associated gene, MFSD2A, is supported by ChIP-seq evidence. g Association between AHR RNA editing levels and MFSD2A expression, assessed using Pearson correlation. h Association between MFSD2A expression and stemness scores, assessed using Pearson correlation. i Comparison of stemness scores between patients with lymph node metastasis (N > 0) and those without (N = 0), evaluated using a t-test RNA editing module 25 contributes to lymph node metastasis. a Network diagram illustrating interactions among genes harboring RNA editing events within Module 25. b Comparison of Module 25 PC1 between patients with lymph node metastasis (N > 0) and those without (N = 0), assessed using a t-test. c Kaplan-Meier survival analysis stratified by the PC1 of Module 25. d Comparison of RNA editing levels of the hub gene AHR between patients with lymph node metastasis (N > 0) and those without (N = 0), evaluated using a t-test. e The transcription factor AHR has 22 manually curated target genes in TRRUST v2, of which 86.36% are annotated as stemness-related genes in BCSCdb. f One AHR-regulated stemness-associated gene, MFSD2A, is supported by ChIP-seq evidence. g Association between AHR RNA editing levels and MFSD2A expression, assessed using Pearson correlation. h Association between MFSD2A expression and stemness scores, assessed using Pearson correlation. i Comparison of stemness scores between patients with lymph node metastasis (N > 0) and those without (N = 0), evaluated using a t-test Discussion This study proposed a comprehensive framework to investigate the interactions among RNA editing events. Previous work has largely focused either on RNA editing enzymes [42] or on individual A-to-I editing events [43, 44]. While enzyme-centric analyses provide a global view of editing activity, they lack site-specific resolution. Conversely, studies cantered on individual editing events overlook the complexity of the RNA editing network within biological systems. Module-based analysis of RNA editing offers a balanced strategy, enabling the characterization of global editing patterns while preserving site-specific functional insights. Several prior studies have applied network-based approaches, such as Weighted Gene Co-expression Network Analysis (WGCNA), to investigate RNA editing [45–47]. WGCNA is effective at identifying broad modules of co-regulated editing events and assessing their associations with clinical phenotypes. In contrast, the Walktrap community detection algorithm employed in this study is specifically designed to identify tightly connected communities, allowing a more precise delineation of intrinsic correlations among RNA editing sites. Given our focus on capturing editing–editing interactions at higher resolution, Walktrap was selected as the primary method for network construction and module detection. The RNA editing module offers valuable insights into gene interactions and potential regulatory mechanisms from the perspective of RNA editing. This complements existing gene interaction networks in databases such as STRING. Notably, strong interactions were identified between the RNA editing of EIF2AK2 and PRKAR2A, which interfere in miRNA-mediated regulation of their expression. This finding reveals a potential functional association between the two genes that is not captured in the STRING database. Furthermore, robust interactions among these genes at both the expression and RNA editing levels were observed across multiple cancer types (Figure S11), providing additional validation for this discovery. To further support the findings, lung cancer cell line datasets before and after ADAR knockdown from GSE93035 were analyzed. The results demonstrated that ADAR knockdown led to decreased expression of EIF2AK2 and PRKAR2A, accompanied by increased expression of MFSD2A, as shown in Figure S12. These observations were highly consistent with the regulatory relationships inferred from the network analysis. Future work will include cell-based and animal-based experiments to further elucidate the biological mechanisms underlying these RNA editing-associated regulatory networks. The high relevance of RNA editing to immune and metastasis is further demonstrated by the RNA editing modules. Specifically, the editing of 12 modules shows a significant correlation with immune infiltration, and the 13 modules are differentially edited in patients with positive lymph node metastasis. Notably, incorporating RNA editing module data yielded prediction results for lymph node metastasis similar to those achieved by adding gene expression data. These analyses highlight that RNA editing modules contain critical information, making them valuable for understanding immune regulation and metastasis. The study of RNA editing modules uncovers the mechanisms through which RNA editing mediates miRNA and transcription factor regulation of gene expression, as well as the role of RBPs in RNA editing. It highlights the close associations and reciprocal regulation between RNA editing and gene expression. Future research will further investigate the interactions between RNA editing and gene expression, aiming to construct a bi-layer editing-gene network to unravel the complex relationships through inter-network interactions. Materials and methods Dataset This study investigated NSCLC patients from TCGA, including both lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) cohorts. Specifically, Specifically, RNA editing data for these patients were obtained from the Synapse database (syn60522234), with the identification pipeline described previously [10]. Gene expression and clinical information were downloaded through the UCSC Xena. The immune infiltration data were assessed using ImmuCellAI methods [25]. Tumor stemness was evaluated based on the mRNA-based Stemness Index (mRNAsi) from the CancerStemnessOnline platform [48]. Univariate and multivariate RNA editing models Two linear regression models were employed to systematically assess the impact of single (univariate) and multiple (multivariate) RNA editing events on gene expression. Model performance was evaluated through 100 repetitions of 10-fold cross-validation, using Mean Squared Error (MSE) and the coefficient of determination (R2). A random seed was set to ensure the reproducibility of all results. In the multivariate model, the editing events were ranked based on the absolute value of the regression coefficients. A paired t-test was used to compare the performance of the univariate and multivariate models. Construction of the RNA editing regulatory network The RNA editing regulatory network was constructed as follows. First, RNA editing events present in more than 20% of samples were retained for downstream analysis to ensure robustness. Pairwise Pearson correlation coefficients were then calculated among all selected editing events to construct an initial weighted network. Edges with an absolute correlation coefficient ≤ 0.2 were removed to reduce spurious associations. To prioritize influential nodes within the network, PageRank scores were computed for each node using an iterative procedure. The initial PageRank value for each node was set to 1/N, where N denotes the total number of nodes. The scores were then updated iteratively according to1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PR(v)=\frac{{1 - d}}{N}+d\sum\limits_{{u \in N(v)}} {\frac{{PR(u) \cdot {w_{u,v}}}}{{\sum\nolimits_{{k \in N(u)}} {{w_{u,k}}} }}} $$\end{document} where d is the damping factor (set to 0.85), N(v) denotes the set of neighbours of node v, and wu, v represents the Pearson correlation coefficient between nodes u and v. Iteration continued until convergence, defined as a change in PageRank scores of less than 1E− 10 between successive iterations. Nodes with PageRank scores above the upper quartile were retained for subsequent analysis. Next, the Walktrap community detection algorithm [49] was applied to identify RNA editing modules, using random walks of length six. The underlying assumption of Walktrap is that nodes belonging to the same community exhibit similar transition probability distributions after t-step random walks. Specifically, two nodes i and j are considered to belong to the same module if their probability distributions over reachable nodes after t steps are similar:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r_{ij}}=\sqrt {\sum\limits_{{k=1}}^{N} {\frac{{{{(P_{{ik}}^{t} - P_{{jk}}^{t})}^2}}}{{d(k)}}} } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{{ik}}^{t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{{jk}}^{t}$$\end{document} denote the probabilities of reaching node k from nodes i and j, respectively, after t steps, d(k) is the degree of node k. During community detection, an iterative recursive partitioning strategy was applied. Modules containing more than 100 nodes were recursively subdivided until all submodules contained fewer than 100 nodes. Finally, modules comprising more than 30 RNA editing events were retained for downstream analyses. Lymph node metastasis prediction models The RNA editing data were incorporated into nine machine learning models to explore its potential in predicting lymph node metastasis. These models included LightGBM [50, 51], LightGBMXT, LightGBMLarge, RandomForestEntr [52], RandomForestGin, ExtraTreesGini [53], ExtraTreesEntr, CatBoost [54], and NeuralNetFastAI [55]. To assess the impact of RNA editing, models using both clinical and RNA editing data were compared to those using only clinical data, with DeLong’s test for statistical comparison. Clinical data included gender, age, histological subtype, and T-stage. RNA editing data were mapped to the top 10 principal components (PCs) for each module and filtered for significant PCs in patients with positive lymph nodes (P < 0.05). Additionally, models incorporating clinical, gene expression, and RNA editing data were compared with models using only clinical and expression data. The expression data were mapped to the top 30 PCs of significant genes in patients with positive lymph nodes (P < 0.05), and the expression PCs were further filtered for significance in these patients (P < 0.05). Each model was trained on a cohort of 950 samples, which were randomly and evenly partitioned into five subsets. Different combinations of these subsets were assigned as the training, validation, and test sets in a 6:2:2 ratio. Model performance was then evaluated on the test sets using the area under the receiver operating characteristic curve (AUC). The mean AUC across all test evaluations was reported. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.