Construction of a diagnostic model for nasopharyngeal carcinoma using a consensus machine learning approach and study of immune infiltration characteristics Authors: Yaozhuang Zhou, Yuanyuan Gu, Chunhua Xie, Yucong Liu, Maosheng Zhang, Liyun Zhang Published: 2026 ABSTRACT AbstractBackgroundNasopharyngeal carcinoma requires reliable diagnostic biomarkers due to its occult location and poor outcomes.MethodsThis study analyzed scRNA-seq data from NPC and control tissues to resolve the tumor microenvironment. CellChat was utilized to infer cell–cell communication. We integrated marker genes from cell clusters, differentially expressed genes (DEGs) from bulk RNA-seq, and key module genes identified by WGCNA to screen candidate genes. Feature selection was then performed using four machine learning algorithms (LASSO, SVM-RFE, Boruta, and XGBoost) to build a robust diagnostic model, and its performance was evaluated with ROC curve analysis. An interactive web application for model visualization was developed using the R Shiny package. We further investigated the prognostic value, immune infiltration association, and functional pathways of the core genes. Potential therapeutic compounds were predicted via the CMAP database and validated by molecular docking.ResultsSingle-cell analysis of 67,535 cells revealed a heterogeneous tumor microenvironment (TME) in NPC, in which all seven identified cell subpopulations made high and balanced contributions to the disease. Four machine learning algorithms consistently screened out four core genes: COL4A2, LAMB1, ACTA2, and CCL2. A diagnostic model based on these genes achieved high accuracy (AUC = 0.933 in the validation set and 0.966 in the external independent validation set). We found that ACTA2 and COL4A2 exhibited strong positive correlations with activated dendritic cells and multiple T cell subsets, whereas CCL2 and LAMB1 showed strong positive correlations with M1 macrophages, neutrophils, and other cell types. Functional enrichment analysis revealed that LAMB1, COL4A2, and ACTA2 primarily drive tumor invasion and remodeling processes such as epithelial-mesenchymal transition and angiogenesis, while CCL2 predominantly governs the activation of the immuno-inflammatory microenvironment. High expression of all four genes was associated with poor prognosis. Computational prediction and molecular docking identified candidate drugs such as parthenolide and panobinostat that can specifically target either CCL2-mediated immuno-inflammatory signaling or the invasion/fibrosis pathways driven by ACTA2 and others, offering a potential strategy for combination therapy targeting the multifaceted pathogenic network in NPC.ConclusionThis study integrated multi-omics data with machine learning to develop a robust four-gene diagnostic model for NPC. The core genes (COL4A2, LAMB1, ACTA2, CCL2) are associated with tumor progression, prognosis, immune regulation, and distinct biological pathways. Our findings provide a valuable tool for the diagnosis and risk stratification of NPC and reveal potential therapeutic targets worthy of further investigation.Keywords: Nasopharyngeal carcinoma, Machine learning, Diagnostic model, Tumor microenvironment Introduction Nasopharyngeal carcinoma (NPC) is a malignancy arising from the nasopharyngeal epithelium of the upper and lateral walls of the nasopharyngeal cavity. It exhibits a strong propensity for metastasis and is one of the most common head and neck tumors. The geographical distribution of NPC is highly distinctive, with a particularly high prevalence in East and Southeast Asia. The incidence rate in regions such as Southern China and Southeast Asia ranges from 15 to 50 cases per 100,000 people, in stark contrast to the incidence rate of approximately 0.4 cases per 100,000 in Western populations [1]. In China, the incidence of nasopharyngeal carcinoma (NPC) also exhibits significant regional disparities, with higher prevalence observed in the southern and southwestern provinces, such as Guangdong, Guangxi, Hainan, Hong Kong, Macau, and Jiangxi [2]. Due to its anatomically concealed location and the frequent nonspecificity of early symptoms, NPC is often difficult to detect early. Consequently, the majority of patients are diagnosed at advanced stages, leading to poor prognosis and a high recurrence rate. Locoregional and/or metastatic recurrence occurs in over 30% of cases, and the overall 5-year survival rate for all stages of NPC is only between 50 and 70% [3]. Therefore, current research efforts should focus on treatment strategies for NPC, exploring novel therapeutic targets and developing new molecularly targeted drugs. Concurrently, there is a need to delve deeper into the underlying molecular mechanisms of NPC and identify new molecular biomarkers. With the rapid advancement of cancer genomics, bulk RNA-seq and scRNA-seq have become primary tools in transcriptomics. An increasing number of studies are utilizing transcriptomics to develop personalized treatment strategies and investigate the underlying molecular mechanisms of cancer. For instance, Mo et al. analyzed the expression profiles of circRNAs in NPC cells with varying metastatic potential. Through in vitro and in vivo functional experiments, they investigated the impact of circPVT1 expression on NPC metastasis, revealing that circPVT1 may serve as a prognostic biomarker or therapeutic target for NPC patients [4]. Ding et al. employed supervised learning to identify Epstein-Barr virus (EBV)-related hub genes and constructed a model to predict NPC prognosis and achieve patient risk stratification [5]. Shan et al. used the ferroptosis pathway as an entry point, applying bioinformatics methods to explore the molecular mechanisms and protein targets for treating cholangiocarcinoma, thereby screening potential therapeutic drugs, which holds significant importance [6]. In the broader field of cancer research, integrating multi-omics data with advanced machine learning algorithms has become a powerful paradigm for identifying novel biomarkers and elucidating tumor heterogeneity. In clear cell renal cell carcinoma (ccRCC), studies have successfully combined machine learning with multi-omics analysis to reveal regulatory T cell (Treg)-related programmed cell death signatures, offering new insights into immunotherapy [7]. Similarly, for bladder cancer, the integration of machine learning models with multi-omics has been used to dissect the prognostic significance of mitotic catastrophe heterogeneity [8], while ADME gene (absorption, distribution, metabolism, excretion)-driven prognostic models have demonstrated breakthrough value in predicting survival and guiding personalized therapy [9]. Cross-disease analytical strategies, such as integrating the genetic crosstalk between atherosclerosis and lung adenocarcinoma, have provided new approaches for precision diagnosis and treatment [10]; moreover, studies dissecting tryptophan metabolism heterogeneity from a multi-omics perspective in ccRCC have highlighted the important role of metabolic remodeling in tumor progression [11]. Recent series of studies in lung adenocarcinoma (LUAD) further underscore the value of integrating multi-omics and machine learning to analyze tumor microenvironment (TME) heterogeneity. Xia et al. [12] identified three molecular subtypes of LUAD with significant survival differences and distinct immune infiltration profiles based on disulfidptosis-related gene expression, and constructed a prognostic model comprising GYS1, NDUFA11, NDUFB10, and SLC7A11; single-cell analysis revealed that these genes are mainly enriched in myeloid cells, suggesting an important role of disulfidptosis-related pathways in modulating the immune landscape of the TME. Gong et al. [13] integrated single-cell transcriptomic data from 153 LUAD patients using machine learning, mapping an immune atlas of over one million cells. They revealed that EGFR-mutant tumors are enriched in TIGIT + regulatory T cells, neutrophils, and macrophages, whereas wild-type tumors exhibit an immune-active TME; they also defined five tumor immune microenvironment (TIME) subtypes, with EGFR-mutant tumors largely clustering into an immunosuppressive subtype associated with poor prognosis. Collectively, these cutting-edge studies demonstrate that machine learning-based multi-omics integration is a reliable approach to unravel complex molecular mechanisms in cancer and to construct robust predictive models. In this study, we analyzed single-cell RNA sequencing (scRNA-seq) data of nasopharyngeal carcinoma to dissect cellular heterogeneity within the tumor microenvironment and employed cell-cell communication analysis to reveal potential interactions among distinct cell subpopulations. Subsequently, we integrated cell signature genes identified from scRNA-seq with DEGs and key modules from WGCNA of bulk RNA-seq data to screen for core candidate genes involved in NPC development. To construct a robust diagnostic model, we systematically applied and compared four machine learning algorithms (LASSO regression, SVM-RFE, Boruta, and XGBoost) for feature selection. Based on the best-performing model, we developed a clinical nomogram for individualized risk prediction and built an interactive web application to facilitate its clinical translation. The diagnostic performance and clinical utility of this model were validated in an independent external dataset and further evaluated by decision curve analysis. In addition, we thoroughly explored the biological and clinical significance of these core genes. Their expression correlations with tumor immune microenvironment infiltration characteristics were analyzed using the CIBERSORT algorithm and ssGSEA. Gene set variation analysis (GSVA) based on the core gene expression profiles revealed the signaling pathways potentially affected by these genes. Finally, we predicted potential therapeutic compounds targeting these genes through the CMAP database, and preliminarily validated the binding feasibility of candidate drugs to their target proteins using molecular docking. This study, through the integration of multi-omics data and computational biology approaches, established a diagnostic model for nasopharyngeal carcinoma with clinical application potential, and systematically explored its underlying biological mechanisms and potential therapeutic strategies. Materials and methods Data acquisition Publicly available single-cell RNA sequencing (scRNA-seq) and bulk RNA sequencing datasets for nasopharyngeal carcinoma (NPC) and non-tumor nasopharyngeal tissues were retrieved from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). The scRNA-seq dataset GSE150430, containing samples from 11 NPC patients and 3 controls, was used for primary analysis of the tumor microenvironment. Multiple bulk RNA-seq datasets (including GSE102349, GSE12452, GSE53819, and GSE13597) were downloaded and merged for model construction and validation. All data processing complied with the data use agreements and ethical guidelines of the original studies. scRNA-seq data processing and analysis scRNA-seq data were processed using the "Seurat" package [14]. Cells in which genes were detected in three or fewer cells, as well as low-quality cells with fewer than 200 detected genes, were excluded from subsequent analysis. The "PercentageFeatureSet" function was used to calculate the percentage of mitochondrial genes, and cells with fewer than 300 or more than 6500 genes expressed, as well as those with a mitochondrial gene percentage exceeding 10%, were removed. The remaining cells were normalized using the "NormalizeData" function. The top 2000 highly variable genes were extracted using the "FindVariableFeatures" function. Principal component analysis was performed on the single-cell samples, and the first 10 principal components were selected for downstream analysis. Cell annotation was performed using the "SingleR" package [15] and the CellMarker database [16]. Differentially expressed genes between different cell types were identified using the "FindAllMarkers" function in the "Seurat" package with parameters min.pct = 0.25 and only.pos = TRUE, and the Wilcoxon rank-sum test was applied to compare gene expression across cell types. Functional enrichment analysis of the marker genes using Gene Ontology (GO) and KEGG [17] pathways was performed with the "ClusterProfiler" package. Cell communication analysis To determine potential interactions between and within cells, the "CellChat" software package was employed [18]. The communication probability of each signaling pathway was inferred by calculating the communication probability of all ligand–receptor pair interactions associated with that pathway. Identification of differentially expressed genes Differential analysis was performed between the 10 normal samples and 31 NPC samples using the "limma" package [19]. Genes with a p-value < 0.05 and an absolute |(LogFC)|> 1 were defined as differentially expressed genes (DEGs). Volcano plots visualizing the DEGs were generated using the "ggplot2" package. Subsequently, the most significantly enriched pathways and biological processes among the DEGs were investigated through Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway and Gene Ontology (GO) enrichment analyses. Construction of weighted gene co-expression network Weighted Gene Co-Expression Network Analysis (WGCNA) was performed to identify clusters (modules) of highly correlated genes, elucidate the association between these gene modules and sample phenotypes (e.g., disease status), and thereby provide insights into potential biological mechanisms [20]. First, data quality was assessed using the goodSamplesGenes function from the "WGCNA" R package to check for missing entries, entries with weights below the threshold, and zero-variance genes, and to return lists of samples and genes meeting the criteria of maximum missing values or low weight values. Subsequently, the optimal soft-thresholding power was selected to construct an adjacency matrix and to achieve an approximate scale-free network topology based on connectivity. Finally, the Pearson correlation coefficient was used to analyze the association between module eigengenes (MEs) and the nasopharyngeal carcinoma phenotype. Construction and validation of the diagnostic model To accurately identify core pathogenic genes, we first integrated marker genes from single-cell clustering analysis, DEGs, and genes from key WGCNA modules using a Venn diagram to obtain intersecting genes. Subsequently, to construct a diagnostic model and screen key features, four different machine learning algorithms were employed for feature selection and model building. LASSO logistic regression is an example of a machine learning technique that defines variables by searching for values corresponding to the lowest possible classification error [21]. Support vector machine recursive feature elimination (SVM-RFE) ranks features and selects the most important classification features [22]; in this study, SVM-RFE was implemented using the e1071 package. The Boruta algorithm determines the most important features in a dataset by comparing the Z-score of each feature with that of its corresponding "shadow features" [23]. XGBoost is a powerful ensemble learning method suitable for tabular datasets. Ensemble learning methods aggregate many individually trained classifiers, and the combined prediction is generally more robust than that of a single classifier [24]. We used each of the above four algorithms to build basic classifiers. To evaluate and compare the performance of the different machine learning algorithms, we calculated the performance metrics of each model as a classifier on the validation set. Finally, the model with the highest area under the receiver operating characteristic curve (AUC) was selected as the optimal diagnostic model. Diagnostic model visualization and interactive web application development A predictive nomogram was constructed using the "rms" R package [25]. To evaluate the clinical utility of the model, Decision Curve Analysis (DCA) was performed. This method calculates the "clinical net benefit" achievable by using the model for intervention decisions across various threshold probabilities and compares it with the simple strategies of "treating all" or "treating none." The analysis was conducted in the R environment using the "rmda" package [26]. To facilitate convenient and real-time clinical application of the model, an interactive web application was developed based on the R Shiny framework [27]. It consists of two main parts: data input and data analysis. The front-end interface allows clinicians to input the expression values of the four key genes for a target patient via sliders or numerical input boxes. The back-end server, embedded with the trained final prediction model, calculates and returns the individualized probability of NPC risk in real-time. The results are visualized as clear numerical percentages and a dynamic dashboard, accompanied by a brief clinical risk stratification interpretation (e.g., "Low Risk," "Moderate Risk," "High Risk"). Survival analysis Based on the expression levels of the key genes, patients were stratified into high-expression and low-expression groups using the median expression value of each gene as the cutoff. Progression-Free Survival (PFS) curves were plotted using the Kaplan–Meier method, and survival differences between groups were compared using the Log-rank test. All statistical analyses were performed in the R language environment utilizing packages such as "survival" and "survminer" for survival analysis. Immune microenvironment analysis In this study, two independent algorithms were employed for immune infiltration analysis to achieve cross-validation, enhance result reliability, and comprehensively characterize the immune infiltration landscape of nasopharyngeal carcinoma (NPC) tissues. The CIBERSORT deconvolution algorithm combined with the LM22 signature matrix was applied to analyze the gene expression profiles of the samples and estimate the relative abundance of 22 immune cell subsets in each sample [28]. Violin plots were generated using the R package ggplot2 to visualize differences in immune cell infiltration levels between NPC patients and normal control samples. Spearman correlation analysis was performed to calculate the correlations between the expression levels of core hub genes and the infiltration abundance of each immune cell subset, thereby exploring the associations between core genes and the tumor immune microenvironment. Simultaneously, a signature gene set of 28 immune cell types was downloaded from the TISIDB database (http://cis.hku.hk/TISIDB/download.php). Single-sample gene set enrichment analysis (ssGSEA) was implemented using the GSVA package in R software to calculate the enrichment score of each immune cell type per sample, quantitatively evaluate the infiltration levels of immune cells in NPC tissues and normal tissues, and independently validate the immune infiltration results [29]. Diagnostic gene enrichment analysis To further investigate how the diagnostic genes contribute to NPC development, patients were categorized based on the expression levels of these genes. The top 30% and bottom 30% of the 88 patients ranked by diagnostic gene expression were defined as the high-expression group and low-expression group, respectively. The Hallmark gene sets were downloaded from the Molecular Signatures Database (MSigDB) [30]. Gene Set Variation Analysis (GSVA) was performed on all genes from samples in both the high-expression and low-expression groups to assess pathway activity. Differences in GSVA enrichment scores between the two groups were analyzed using the "limma" R package [29]. Small molecule drug prediction and molecular docking validation Potential small-molecule drugs for treating nasopharyngeal carcinoma (NPC) were screened using the CMAP database [31]. Candidate genes were separately compiled into "upregulated" and "downregulated" gene lists and submitted to the database. A connectivity score closer to -100 indicates a higher therapeutic potential for the compound. In this study, a connectivity score of less than -80 was used as the threshold to filter candidate small-molecule compounds. To investigate the binding mechanisms between the candidate compounds and their potential target proteins, the three-dimensional structures of the key target proteins were retrieved from the RCSB Protein Data Bank (PDB, https://www.rcsb.org/). The target proteins and their corresponding PDB accession codes were as follows: CCL2 (4DN3), COL4A2 (5NB2), LAMB1 (5XAU), and ACTA2 (AF_AFP62736F1). The three-dimensional structures of the candidate compounds were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/), with the following PubChem Compound Identifiers (CIDs): Panobinostat (6918837), Droxinostat (568416), Parthenolide (6473881), Tipifarnib (159324), Nocodazole (4122), and Sulforaphane (5350). Molecular docking was performed using AutoDock software. A binding energy of < 0 kcal/mol indicated spontaneous binding between the ligand and the receptor. Visualization of selected docking results was conducted using PyMOL software. Statistical analysis The statistical analyses employed in this study were conducted using R software (version 4.4.1). A p-value < 0.05 or an adjusted p-value < 0.05 was considered statistically significant. Results NPC single-cell atlas We collected single-cell RNA sequencing data from 14 independent samples, including 11 nasopharyngeal carcinoma (NPC) samples and 3 nasopharyngeal lymphoid hyperplasia (NLH) samples, initially obtaining gene expression profiles of 73,763 single cells. To further ensure data quality, we implemented a rigorous quality control pipeline, ultimately retaining 67,535 high-quality cells for subsequent analysis (Fig. 1A). After data normalization, principal component analysis was performed based on the expression of 2,000 highly variable genes (Fig. 1B), and all cells were clustered according to the significant principal components, resulting in the identification of 17 cell clusters (Fig. 1C). Dimensionality reduction and visualization were performed using t-SNE, and cells were annotated based on marker genes of known cell types, leading to the successful identification of seven major cell types (Fig. 1E). Figure 1D shows the t-SNE expression distribution of cell type-specific marker genes. These cell types included B cells, plasma cells, T cells, monocytes, macrophages, epithelial cells, and fibroblasts. Analysis of cell type proportions in tumor versus control tissues (Fig. 1F) revealed that while T cells retained a certain proportion in NPC tissues, the proportion of B cells was significantly reduced, suggesting that B cell infiltration in the tumor microenvironment may be suppressed or excluded. Meanwhile, the proportions of macrophages and monocytes in NPC tissues were slightly higher than those in NLH tissues, indicating enhanced infiltration of myeloid immune cells into the tumor microenvironment, which may be associated with the immunomodulatory functions of tumor-associated macrophages (TAMs). Collectively, these results reveal the heterogeneous characteristics of the cellular composition of the nasopharyngeal carcinoma tumor microenvironment. Single-cell atlas of nasopharyngeal carcinoma. A Quality control and cell filtering of the single-cell data. B Principal component analysis based on highly variable genes. C Unsupervised clustering analysis of cells. D Heatmap of representative marker gene expression for each cell cluster. E t-SNE visualization and cell type annotation. F Analysis of differences in cell type proportions between NPC and control tissues Contribution and functional enrichment analysis of different cell subpopulations to NPC By screening the top 100 genes with high performance in both the control and disease groups, we quantified the differences in expression levels and expression proportions of these genes within each cellular subtype. To assess their contribution to the disease, we ultimately used the square root of FC × PctProp as an indicator of disease contribution. We found that all seven cellular subtypes exhibited high and balanced contributions to NPC (Fig. 2A). This suggests that NPC is not dominated by a single cell type but is likely driven by a complex network involving multiple interacting cell types. Therefore, all seven cell types were included in subsequent analyses. Characterization and functional analysis of cell types in the nasopharyngeal carcinoma microenvironment. A Disease contribution of cell types based on the FC × PctProp metric. B KEGG pathway enrichment analysis for each cellular subtype in the NPC microenvironment. C–E GO functional enrichment analysis for each cellular subtype in the NPC microenvironment Using 1899 marker genes identified via FindAllMarkers and the Wilcoxon test, we performed functional enrichment analysis on the seven cell types in the NPC microenvironment. The GO analysis results (Fig. 2C–E) showed that immune cells were significantly enriched in immune-related pathways, including antigen presentation, immune response regulation, and interferon signaling, reflecting their active immune surveillance functions. In contrast, epithelial cells and fibroblasts were primarily involved in structural processes such as cell junctions, adhesion, and extracellular matrix organization. Notably, multiple cell types were commonly enriched in pathways related to RNA splicing, cytoplasmic translation, and ribosome biogenesis, suggesting extensive post-transcriptional regulation and a highly active state of protein synthesis within the tumor microenvironment. KEGG analysis further confirmed the specific distribution of immune-related pathways (e.g., B cell receptor signaling, antigen processing and presentation) and structure-related pathways (e.g., focal adhesion, ECM-receptor interaction) (Fig. 2B). In summary, the nasopharyngeal carcinoma microenvironment constitutes a functionally synergistic ecosystem, where immune cells dominate immune responses, structural cells maintain the microenvironmental architecture, and the system as a whole exhibits vigorous metabolism and active protein synthesis. NPCs have different intercellular communication modes Through the analysis of intercellular communication among cell subpopulations within the tumor microenvironment (TME), we systematically mapped ligand-receptor interactions across different cell types and inferred potential intercellular interaction networks. Analysis of both the quantity and intensity of cell interactions revealed extensive communication connections among immune cells, as well as between immune cells and stromal cells (Fig. 3A). Among these interactions, those between plasma cells and T cells, as well as between macrophages and B cells, were particularly prominent. Meanwhile, monocytes and fibroblasts, as well as epithelial cells and other cell types, also exhibited active signaling crosstalk. Ligand-receptor interaction analysis further unveiled a pattern of multi-pathway synergistic regulation. Macrophages, acting as key signal senders, widely transmit signals to B cells, T cells, and epithelial cells, involving multiple signaling pathways such as LGALS9 → CD44/CD45 and MIF (Fig. 3B). In terms of outgoing signaling patterns, macrophages and fibroblasts exhibited strong signal output capabilities across multiple pathways, including GALECTIN, TNF, SPP1, and CXCL, while plasma cells and B cells showed relatively weaker signaling activity (Fig. 3C). Conversely, analysis of incoming signaling patterns revealed that B cells and T cells were the primary receivers of multiple signaling pathways, actively accepting signal inputs from pathways such as TNF, CXCL, GALECTIN, and SPP1 (Fig. 3D). Cell Communication Analysis. A Analysis of intercellular interaction networks. B Key ligand–receptor mediated intercellular communication. C Outgoing signaling patterns of major cell types. D Incoming signaling patterns of major cell types Identification and functional enrichment analysis of DEGs in bulk RNA-seq data A total of 908 DEGs were identified, comprising 356 upregulated and 552 downregulated genes (Fig. 4A). GO analysis indicated that the DEGs were primarily enriched in cell cycle-related functions, including mitotic sister chromatid segregation, chromosome segregation, mitotic nuclear division, regulation of chromosome segregation, and regulation of DNA replication (Fig. 4C–E). KEGG enrichment results revealed that the DEGs were significantly enriched in pathways such as the cell cycle, viral protein interaction with cytokine and cytokine receptor, IL-17 signaling pathway, ECM-receptor interaction, cytokine-cytokine receptor interaction, and DNA replication (Fig. 4B). Identification of DEGs between NPC samples and normal samples. A Volcano plot of DEGs, with red dots representing upregulated genes and blue dots representing downregulated genes. B–E KEGG and GO enrichment analyses of the DEGs. Identification of co-expression gene modules in nasopharyngeal carcinoma WGCNA was used to identify gene modules involved in the development and progression of NPC. First, the dataset samples were divided into two groups (NPC and normal), and no low-quality samples or gene names were observed. During co-expression network construction, when the scale-free topology fit index exceeded 0.85, the soft-thresholding power β was set to 9 (Fig. 5A). Through hierarchical clustering analysis and dynamic branch cutting of the gene dendrogram, the genes were classified into 12 modules (Fig. 5B, C). The MEblue module showed a significant correlation with NPC (r = 0.72, p = 1e-07) and was selected as the key module (Fig. 5D). This module contained 766 genes closely associated with NPC. GO and KEGG enrichment analyses were performed on the core module. GO analysis revealed that the differentially expressed genes were mainly enriched in biological processes such as chromosome segregation, mitotic nuclear division, and the cell cycle (Fig. 5E). KEGG pathway analysis further indicated that these genes were significantly involved in the cell cycle, p53 signaling pathway, DNA replication, and cancer-related pathways (Fig. 5F). WGCNA combined with enrichment analysis successfully identified NPC-associated key modules and functional pathways, providing core clues for a deeper understanding of the molecular mechanisms of NPC. Identification of co-expression gene modules associated with nasopharyngeal carcinoma based on WGCNA. A Determination of the soft threshold and scale-free topology fitting. B, C Identification and cluster analysis of gene co-expression modules. D Scatter plot showing the correlation between the key module (MEblue) and the NPC phenotype. E GO analysis of the key module. F KEGG analysis of the key module Construction of a diagnostic model By taking the intersection of marker genes from cell clusters, DEGs, and key module genes using a Venn diagram, a total of 31 overlapping genes were obtained and defined as candidate genes (Fig. 6A). Subsequently, four different machine learning methods (XGBoost, Lasso, SVM-RFE, and Boruta) were used to select key genes and construct a diagnostic model. In terms of feature importance, XGBoost selected 4 genes, Lasso selected 6 genes, SVM-RFE selected 11 genes, and Boruta selected 22 genes. The genes selected by all machine learning algorithms, i.e., the key genes (COL4A2, LAMB1, ACTA2, and CCL2), were further used to construct the diagnostic model (Fig. 6B). The Wilcoxon rank-sum test was used to compare the expression levels of the key genes between the NPC and control groups, and all four genes were found to be upregulated in NPC (Fig. 6C, D). Furthermore, patients were divided into high- and low-expression groups based on the median expression of each key gene, and survival curves were plotted using the Kaplan-Meier method, with the log-rank test used to compare survival differences between groups. The results showed that patients with high expression of COL4A2, LAMB1, ACTA2, and CCL2 had significantly shorter progression-free survival (PFS) than those with low expression (p < 0.05) (Fig. 6E–H). High expression of the key genes was closely associated with poor prognosis in NPC patients. Screening, identification, and validation of diagnostic genes for nasopharyngeal carcinoma. A Venn diagram showing the intersection of cell cluster marker genes, DEGs, and key module genes. B Feature importance and intersection of key genes screened by four machine learning methods. C, D Expression differences of key genes between the NPC and control groups. E–H Kaplan–Meier curves for progression-free survival (PFS) of patients in the high- and low-expression groups of key genes Validation of the diagnostic model for nasopharyngeal carcinoma using external datasets. To compare the performance of each machine learning algorithm, we evaluated the performance of each model as a classifier on the validation set. As shown in Table 1, the Lasso model achieved the highest AUC (0.867). The XGBoost model had the lowest AUC (0.707). The number of genes selected by each model also differed, with the Boruta model selecting the most genes and the XGBoost model selecting the fewest. Furthermore, we further focused on the overlapping genes selected by the four machine learning algorithms and evaluated the diagnostic performance of each individual overlapping gene in nasopharyngeal carcinoma. The results showed that LAMB1 performed best, with an AUC value reaching 0.92 (Fig. 7A). Subsequently, the genes selected by all models were combined to construct a diagnostic model, which achieved an AUC value of 0.933, indicating high predictive value (Fig. 7B). The predictive performance of the diagnostic model was then evaluated using the external validation set GSE53819, and the results showed an AUC of 0.966 for the diagnostic model (Fig. 7C). These results confirm that the diagnostic model constructed with the genes (COL4A2, LAMB1, ACTA2, and CCL2) has excellent diagnostic efficiency. Therefore, if certain features are determined to be associated with the diagnosis of nasopharyngeal carcinoma, this may in the future prompt investigations into therapeutic interventions targeting specific genes. Table 1Performance of the four classifiers in the validation setXGBoostLassoBorutaSVM-RFEGenes selected by model, n462211Sensitivity0.640.600.600.64Specificity1111Accuracy0.680.650.630.68Balanced_accuracy0.820.80.80.82AUC0.7070.8670.840.747 Performance of the four classifiers in the validation set Construction and validation of the diagnostic model for nasopharyngeal carcinoma. A Evaluation of the independent diagnostic performance of overlapping key genes. B Performance of the integrated diagnostic model based on key genes in the validation set. C Validation of the diagnostic model in an external independent cohort Visualization of the diagnostic model To facilitate the clinical application of this diagnostic model, we developed a visual prediction tool based on a nomogram (Fig. 8A). This nomogram integrates the expression levels of four genes, allowing for an intuitive calculation of an individual's total score and corresponding disease probability by simply summing the points assigned to each variable. Decision curve analysis demonstrated that across a wide range of probability thresholds (if the threshold probability for patients or clinicians is > 10%), using this gene-based diagnostic model for decision-making yields significant clinical net benefit, indicating the model's strong clinical utility (Fig. 8B). Furthermore, we created an interactive web application (built on the R Shiny framework), enabling clinicians to conveniently input patients' gene expression data and obtain real-time, personalized risk predictions (Fig. 8C). Development of a visual predictive tool for the gene diagnostic model. A Risk prediction nomogram based on the expression of four genes. B Decision curve analysis assessing the model's clinical net benefit. C Interface of the interactive risk prediction web tool based on R Shiny Analyzing the correlation between key genes and immune cell infiltration in NPC To comprehensively characterize the immune infiltration landscape of nasopharyngeal carcinoma (NPC), this study employed two independent algorithms, CIBERSORT and ssGSEA, for cross-validation (Fig. 9A, B). The results showed that, compared with normal tissues, M1 macrophages, activated dendritic cells, neutrophils, and regulatory T cells (Tregs) were significantly elevated in NPC tissues according to both algorithms, whereas naïve B cells and resting NK cells were significantly reduced, confirming that the NPC tumor microenvironment is characterized by chronic inflammatory activation, adaptive immune imbalance, and immune evasion. Correlation analysis revealed that ACTA2 and COL4A2 were mainly significantly positively correlated with activated dendritic cells, activated CD4⁺ memory T cells, and Tregs (Fig. 9C, D); whereas CCL2 and LAMB1 were closely positively correlated with the infiltration levels of M1 macrophages and neutrophils (Fig. 9E, F). These associations reveal the intrinsic links between the core hub genes and the tumor immune infiltration characteristics, preliminarily elucidate the potential key molecules driving the formation of the local inflammatory and immunosuppressive microenvironment in NPC, and provide a reliable theoretical basis for in-depth investigation of the immune mechanisms of NPC and the development of novel combination immunotherapy strategies. Immune infiltration characteristics of nasopharyngeal carcinoma and normal tissues and immune correlation analysis of core genes. A Differences in infiltration of 22 immune cell types between NPC and normal tissues analyzed by the CIBERSORT algorithm (violin plot). B Infiltration enrichment characteristics of 28 immune cell types validated by the ssGSEA algorithm (boxplot). C Spearman correlation between ACTA2 expression level and infiltration of key immune cells. D Spearman correlation between COL4A2 expression level and infiltration of key immune cells. E Spearman correlation between CCL2 expression level and infiltration of key immune cells. F Spearman correlation between LAMB1 expression level and infiltration of key immune cells Functional analysis of diagnostic genes To gain deeper insights into the key pathogenic mechanisms of nasopharyngeal carcinoma and explore potential therapeutic targets, we performed functional enrichment analysis of four diagnostic-related genes (LAMB1, COL4A2, ACTA2, and CCL2). Based on 88 NPC samples, the top 30% and bottom 30% of expression levels for each gene were defined as high- and low-expression groups, respectively, and GSVA was conducted. High expression of LAMB1 was significantly associated with the activation of pro-tumorigenic pathways such as epithelial-mesenchymal transition (EMT), angiogenesis, glycolysis, and WNT/β-catenin signaling. Moreover, high LAMB1 expression was also correlated with upregulation of the hypoxia signaling pathway, indicating its involvement in tumor hypoxic adaptation. Notably, high LAMB1 expression was associated with downregulation of immune-related pathways, including interferon-γ response and the complement system (Fig. 10A). High expression of COL4A2 was also closely related to the upregulation of EMT, Hedgehog signaling, angiogenesis, Notch, and WNT/β-catenin signaling, suggesting that this gene may drive NPC progression through the activation of multiple developmental and pro-invasive pathways. In addition, activation of the TNFα signaling pathway (via NFκB) implies that COL4A2 is also involved in regulating tumor-associated inflammation and the immune microenvironment. Conversely, significant downregulation of metabolic pathways such as bile acid metabolism represents a core feature of metabolic reprogramming in NPC tumor cells (Fig. 10B). High expression of ACTA2 was significantly enriched in EMT, angiogenesis, and myogenesis pathways (Fig. 10C). In contrast to the three genes above, CCL2 exhibited a distinct pattern of enrichment characterized by immuno-inflammatory features. Its high expression was associated with strong activation of TNFα/NFκB signaling, inflammatory response, interferon (α/γ) response, IL-6/JAK/STAT3 signaling, and the complement system, highlighting the core role of CCL2 as a key chemokine in recruiting and activating immune cells and shaping the inflammatory tumor microenvironment. The upregulation of apoptosis and p53 pathways may reflect the cellular stress state induced by this microenvironment (Fig. 10D). In summary, LAMB1, COL4A2, and ACTA2 primarily promote tumor invasion and stromal remodeling by directly regulating EMT, angiogenesis, and development-related signaling pathways, whereas CCL2 mainly influences the tumor microenvironment through activating a broad immuno-inflammatory network. The differential expression of these four genes collectively constitutes a complex and multifaceted pathogenic network in NPC, providing potential targets for combination therapeutic strategies directed at distinct mechanistic layers. Diagnostic gene signatures reveal distinct functional modules in nasopharyngeal carcinoma through GSVA enrichment analysis. A LAMB1 associates with epithelial-mesenchymal transition, angiogenesis, and immunosuppression. B COL4A2 upregulates development-related signaling pathways and inflammatory responses. C ACTA2 links stromal activation and mesenchymal transition. D CCL2 dominates the pro-tumor immunoinflammatory microenvironment Docking of small molecule compounds To translate the pathogenic mechanisms of the above diagnostic genes into therapeutic strategies, we performed reverse drug screening using the CMAP database to identify small-molecule compounds that may reverse the gene signature expression profile. Using a connectivity score < − 80 as the threshold, we screened a set of candidate drugs with potential therapeutic value (Table 1). These drugs cover multiple anti-tumor mechanisms, including epigenetic regulation, signaling pathway inhibition, cytoskeleton targeting, and metabolic intervention. Further molecular docking simulations revealed that these drugs exhibit good binding affinities (docking score < − 5.0 kJ/mol) with the proteins encoded by the corresponding diagnostic genes, suggesting the possibility of direct interaction and providing a structural basis for subsequent mechanistic studies (Table 2). Table 2Small molecule drugsDrug nameCategoryCMAP connectivity scoreDocked geneBinding energy (kJ/mol)PanobinostatHDAC inhibitor− 98.45CCL2− 7.908DroxinostatHDAC inhibitor− 99.19ACTA2− 9.749ParthenolideNF-κB pathway inhibitor− 99.33CCL2/LAMB1− 12.008/− 16.568TipifarnibFarnesyltransferase inhibitor− 97.48ACTA2− 19.707NocodazoleTubulin inhibitor− 97.81ACTA2− 18.451SulforaphaneAntineoplastic agent− 99.44COL4A2− 10.544 Small molecule drugs The predicted drugs parthenolide and panobinostat are potentially associated with the intense immune-inflammatory signaling enriched by CCL2. Parthenolide, a natural NF-κB inhibitor, has been shown to block TNFα-induced NF-κB activation [32]; therefore, it is speculated that parthenolide may indirectly regulate CCL2 expression and CCL2-mediated macrophage recruitment through inhibition of the NF-κB pathway, offering a potential strategy to reverse the CCL2-driven pro-tumor inflammatory microenvironment. Previous studies have shown that sulforaphane promotes the transcription of tumor suppressor proteins and exhibits anticancer and anti-inflammatory activities [33]. According to our GSVA results, ACTA2 is significantly enriched in EMT, angiogenesis, and myogenesis pathways, while tipifarnib and nocodazole have been reported to target and inhibit tumor-associated fibrosis and cell motility, aligning closely with the functional characteristics of ACTA2 in the NPC tumor microenvironment. To further elucidate the potential binding mechanisms between the candidate compounds and the four core target proteins, we performed a detailed analysis of the non-covalent interactions from the molecular docking results (Fig. 11). The molecular docking simulations showed that all candidate compounds could stably bind to the active pockets of their target proteins through multiple non-covalent interactions. Among these, hydrogen bonds were the core driving force for binding stability. Key binding residues included MET-143 of COL4A2, TYR-2881 of LAMB1, LYS-20/GLY-17 of ACTA2, and ARG-24/PHE-72 of CCL2, all of which formed strong hydrogen bonds with the corresponding ligands (bond lengths mostly < 3.0 Å). Hydrophobic interactions mediated by non-polar residues such as Leu, Met, and Phe further anchored the ligands within the binding pocket, enhancing binding affinity. π-π stacking and cation-π interactions provided additional binding stability for ligands containing aromatic rings. These molecular docking results preliminarily confirm the high binding affinity and specific binding modes between the candidate compounds and the target proteins, providing a theoretical reference for subsequent investigations into their potential anti-NPC effects. Molecular docking of small-molecule drugs with diagnostic gene proteins. A Molecular docking results of COL4A2 with sulforaphane. B Molecular docking results of LAMB1 with Parthenolide. C Molecular docking results of ACTA2 with Droxinostat. D Molecular docking results of CCL2 with Panobinostat. E Molecular docking results of ACTA2 with Nocodazole. F Molecular docking results of ACTA2 with Tipifarnib Discussion NPC is a malignant head and neck tumor closely associated with Epstein-Barr virus (EBV) infection, and its development and progression are profoundly influenced by the complex interactions within the tumor microenvironment (TME). This study integrates single-cell RNA sequencing (scRNA-seq) and bulk RNA sequencing data to systematically characterize the cellular composition, functional states, and interaction networks of the NPC TME. Building on this foundation, a multi-gene-based diagnostic model was constructed and validated, providing critical insights for a deeper understanding of the biological features of NPC and the development of novel diagnostic and therapeutic strategies. First, our single-cell atlas revealed the high heterogeneity of the NPC TME. Compared with normal nasopharyngeal lymphoid hyperplasia (NLH) tissues, the proportion of T cells in NPC tissues remained relatively stable, whereas the proportion of B cells was significantly decreased. This contrasts with the trend reported in several studies showing a positive correlation between B cell infiltration and prognosis in various solid tumors [34], suggesting a unique B cell exclusion effect or a regional immunosuppressive microenvironment in the NPC TME. In conjunction with our cell-cell communication analysis, these results further confirmed that B cells act as central signal receivers in the global ligand-receptor interaction network, serving as key responder cells to exogenous regulatory signals within the tumor microenvironment. The marked reduction in their overall abundance not only reflects a systemic weakening of the humoral antitumor immune response; accumulating evidence also indicates that the tumor microenvironment can induce functional exhaustion of infiltrating B cells and promote their polarization toward regulatory B cells (Bregs), which amplify local immunosuppression and mediate tumor immune escape through the secretion of anti-inflammatory cytokines such as IL-10 and TGF-β [35, 36]. Based on the signaling interaction features of this study, we speculate that the reduced B-cell population in NPC does not represent an antitumor activated subtype; rather, the residual infiltrating B cells are most likely in a functionally suppressed state and tend to differentiate toward regulatory B cells (Bregs), thereby synergistically promoting the malignant progression of nasopharyngeal carcinoma. Meanwhile, this study revealed a significant increase in the infiltration proportion of monocyte/macrophage populations. Functional enrichment analysis confirmed that immune cells were mainly enriched in pathways related to antigen presentation and immune regulation, whereas epithelial cells and fibroblasts were preferentially involved in tissue structure maintenance and extracellular matrix (ECM) remodeling. Almost all cell subtypes were enriched in ribosome biogenesis and protein synthesis pathways, collectively corroborating the overall hypermetabolic and hyperactive biosynthetic characteristics of the NPC TME, thereby providing an important metabolic basis for rapid tumor proliferation and poor prognosis. Cell communication analysis further elucidated a complex signaling regulatory network cooperatively constructed by multiple cell types. Macrophages and fibroblasts served as the main signal senders, actively reshaping the immune microenvironment by secreting key ligands such as GALECTIN, TNF, SPP1, and CXCL. In contrast, B/T cells acted as core signal receivers, continuously receiving inhibitory regulation from upstream myeloid and stromal cells. Notably, interaction pairs such as LGALS9-CD44/CD45 and MIF were significantly enriched, providing potential intervention targets for subsequently reversing B cell inhibition and blocking aberrant cell-cell crosstalk. Based on multi-omics data integration and machine learning, we successfully screened and validated a diagnostic model consisting of four key genes (COL4A2, LAMB1, ACTA2, and CCL2). In external validation cohorts, the AUC values were all greater than 0.9, indicating high diagnostic value. We further constructed a visualized nomogram and an interactive web tool to facilitate the future clinical application of this model. Decision curve analysis confirmed that the model provides a positive clinical net benefit over a wide range of threshold probabilities. COL4A2 is a key component of the basement membrane and ECM, and its overexpression is closely associated with tumor invasion, metastasis, and poor prognosis [37]. Consistent with our findings, high expression of COL4A2 was significantly associated with shorter progression-free survival (PFS) in patients. The marked downregulation of pathways such as bile acid metabolism in the COL4A2-high expression group represents an important manifestation of metabolic reprogramming in nasopharyngeal carcinoma tumor cells. Bile acid metabolism is a characteristic functional pathway of normal nasopharyngeal epithelial cells; its downregulation reflects the dedifferentiated state of tumor cells, in which they lose the metabolic features of normal epithelial cells and shift toward the glycolysis-centered Warburg effect, thereby providing energy and biosynthetic substrates for rapid proliferation [38]. Moreover, bile acids and their metabolites (e.g., secondary bile acids) exert anti-tumor effects by inhibiting tumor cell proliferation and modulating the immune microenvironment. Downregulation of this pathway relieves the bile acid-mediated suppression of tumor growth, thereby creating favorable conditions for the progression of nasopharyngeal carcinoma [39–41]. Furthermore, the downregulation of bile acid metabolism is highly synergistic with the upregulation of pro-invasive pathways such as EMT and WNT/β-catenin, suggesting that COL4A2 may drive invasion and metastasis in nasopharyngeal carcinoma by regulating cellular metabolic reprogramming and disrupting normal epithelial cell function, further elucidating the key role of COL4A2 in the pathogenic network of NPC. Upregulation of LAMB1 has been implicated in various cancers, including hepatocellular carcinoma [42], prostate cancer [43], and glioma [44]. Our study revealed that high LAMB1 expression in nasopharyngeal carcinoma is associated with the activation of multiple pro-tumorigenic pathways; moreover, high LAMB1 expression is also correlated with upregulation of the hypoxia signaling pathway, indicating its involvement in tumor hypoxic adaptation. ACTA2 (α-SMA) is a classic marker of cancer-associated fibroblast (CAF) activation, and CAFs promote tumor progression by remodeling the ECM and secreting factors [45]. CCL2 is a key chemokine that recruits monocytes/macrophages to tumor sites, serving as an important bridge linking inflammation and cancer [46]. Further immune infiltration analysis provided a cellular-level explanation for this molecular network. ACTA2 and COL4A2 were positively correlated with regulatory T cells (Tregs) and tolerogenic dendritic cells, suggesting that activated stroma may influence antigen presentation and T-cell function [47]. In contrast, CCL2 and LAMB1 showed strong correlations with myeloid cell infiltration, including M1 macrophages and neutrophils, directly confirming the central role of CCL2 in shaping the immunosuppressive myeloid environment. Consistent with our observations, Zhang et al. recently reported in lung adenocarcinoma that metabolic reprogramming features can drive immune exclusion and a “cold” tumor microenvironment, in which high expression of the key metabolic gene SLC25A1 is significantly associated with spatial exclusion of CD8⁺ T cells and reduced anti-PD-1 efficacy, whereas SLC25A1 knockout enhances CD8⁺ T-cell infiltration and cytotoxic function [48]. These findings from a metabolic perspective support the biological significance of the associations between stromal/chemokine genes (e.g., CCL2, ACTA2) and the infiltration intensity of specific immune subsets in our study, further reinforcing the generalizability of the “molecular signature-immune exclusion-clinical outcome” regulatory axis. Therefore, these four genes collectively constitute a core axis driving the vicious cycle of “stromal remodeling-immune recruitment” within the NPC microenvironment. Based on the above molecular features, we predicted potential small-molecule drugs that may reverse the NPC gene expression profile through the CMAP database, including HDAC inhibitors (e.g., panobinostat), EGFR inhibitors (e.g., neratinib), and NF-κB pathway inhibitors (e.g., parthenolide). It should be emphasized that these predictions are purely computational, providing only theoretical clues regarding candidate compounds. Their actual anti-NPC activity and regulatory relationship with the four key genes must be verified through subsequent in vitro/in vivo experiments. This study still has several limitations that need to be addressed in future work. First, the mechanistic inferences in this study are primarily based on GSVA pathway enrichment analysis and gene expression correlation analysis, and the molecular docking results are only structure-based theoretical simulations, lacking direct functional evidence from in vitro cellular experiments or in vivo animal models. Second, the core analysis of this study focuses on transcriptome-level gene expression data, and the expression differences of the four key prognostic genes have not yet been independently validated at the protein level. Meanwhile, the candidate small-molecule drugs predicted through the CMAP database remain at the computational simulation stage, and their target specificity, anti-tumor activity, and clinical translational value await further experimental validation. In subsequent studies, multi-level expression validation will be carried out using techniques such as qRT-PCR and immunohistochemistry, and the biological functions and drug intervention potential of the key genes will be systematically elucidated through cellular phenotypic assays, organoid models, and animal experiments, thereby providing comprehensive experimental evidence for the reliability of the study's conclusions. In the future, in-depth mechanistic dissection and targeted drug development centered on these key molecules hold promise for providing new strategies and theoretical foundations for precision treatment and prognosis improvement of nasopharyngeal carcinoma. Author contributions Y.Z. and M.Z. conceived and designed the study. Y.Z. performed the experiments and wrote the initial draft of the manuscript. Y.G. and C.X. conducted the data analysis and visualization. Y.L. performed critical review and revision of the research methodology and results analysis. M.Z. and L.Z. provided essential research resources and technical guidance. All authors (Y.Z., Y.G., C.X., Y.L., M.Z., L.Z.) participated in reviewing, discussing, and finalizing the manuscript. Funding This research was supported by the Guangxi Natural Science Foundation under Grant No.2024JJA140662. Data availability The datasets analyzed during this study are available in the following public repositories: Single-cell RNA sequencing (scRNA-seq) and bulk RNA-seq data were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) under accession numbers GSE150430, GSE102349, GSE12452, GSE53819, and GSE13597.Cell type annotation was performed using the CellMarker database (http://bio-bigdata.hrbmu.edu.cn/CellMarker/). Gene set enrichment analysis employed the Hallmark gene sets from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb). Drug prediction was performed using the Connectivity Map (CMAP) database (https://clue.io/). Molecular docking studies utilized protein structures from the Protein Data Bank (PDB) (https://www.rcsb.org/) and ligand information from PubChem (https://pubchem.ncbi.nlm.nih.gov/). The specific accession codes for PDB structures and PubChem Compound IDs are provided in the Methods section. The code and analytical scripts used in this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This study is based on the analysis of publicly available datasets and does not involve any new experiments on human participants or animals by the authors. Therefore, ethical approval is not applicable to this manuscript. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Maosheng Zhang, Email: zmsinfo@gxmu.edu.cn. Liyun Zhang, Email: 219510@sr.gxmu.edu.cn. References Associated Data Data Availability Statement The datasets analyzed during this study are available in the following public repositories: Single-cell RNA sequencing (scRNA-seq) and bulk RNA-seq data were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) under accession numbers GSE150430, GSE102349, GSE12452, GSE53819, and GSE13597.Cell type annotation was performed using the CellMarker database (http://bio-bigdata.hrbmu.edu.cn/CellMarker/). Gene set enrichment analysis employed the Hallmark gene sets from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb). Drug prediction was performed using the Connectivity Map (CMAP) database (https://clue.io/). Molecular docking studies utilized protein structures from the Protein Data Bank (PDB) (https://www.rcsb.org/) and ligand information from PubChem (https://pubchem.ncbi.nlm.nih.gov/). The specific accession codes for PDB structures and PubChem Compound IDs are provided in the Methods section. The code and analytical scripts used in this study are available from the corresponding author upon reasonable request.