Unveiling the Diagnostic Value and Potential Therapeutic Targets of Phenylalanine Metabolism in Pancreatic Cancer via Integrated Multi‐Omics and Machine Learning Authors: Xing Liu, Yibin Li, Fangfang Qin, Jianghong Ou Published: 2026 ABSTRACT ABSTRACTPancreatic cancer (PC) presents a significant global health challenge because of its high mortality rate, highlighting the urgent requirement for effective early diagnostic and therapeutic strategies. This study examined the function of phenylalanine metabolism in PC and developed a high‐accuracy diagnostic model by integrating metabolomics, Mendelian randomization (MR), and machine learning (ML) algorithms. Initially, MR analysis was conducted on 55 plasma metabolites, revealing a significant causal link between phenylalanine and PC. Utilizing GeneCards and public transcriptomic databases, we determined eight differentially expressed genes (DEGs) in PC associated with phenylalanine. Based on these genes, we utilized 12 ML algorithms, totaling 113 combinations, to select the optimal diagnostic model. We applied Shapley Additive exPlanations (SHAP) for feature interpretation and constructed a prognostic nomogram with strong predictive performance by incorporating clinical variables. Furthermore, immune infiltration analysis demonstrated strong connections between these key genes and specific immune cell populations. Based on the SHAP value, we conducted single‐cell RNA sequencing (scRNA‐seq) data and simulated gene knockout analyses using SLC6A14 as the key gene. Drug target prediction‐guided molecular docking and molecular dynamics simulations, focusing on the core gene SLC6A14, confirmed the high binding stability of candidate compounds. Finally, in vitro cell experiments quantitative real‐time PCR (RT‐qPCR) verified the expression trends of the key genes in PC cell lines. In conclusion, this study successfully developed an ML diagnostic model with high biological interpretability. This analysis aims to identify biomarkers related to phenylalanine metabolism and potential therapeutic drugs for PC, offering new strategies for personalized targeted therapy of PC.Keywords: diagnostic markers, machine learning, mendelian randomization, metabolomics, pancreatic cancer Abbreviations body mass index decision curve analysis differentially expressed genes false discovery rate gene expression omnibus gene set enrichment analysis gene set variation analysis Kyoto encyclopedia of genes and genomes linear discriminant analysis machine learning Mendelian randomization orthogonal partial least squares discriminant analysis pancreatic cancer principal component analysis shapley additive explanations tumor immune microenvironment Introduction Pancreatic cancer (PC) is among the most lethal malignancies worldwide, with both its occurrence and mortality rates having risen significantly over recent decades. Data from the Global Burden of Disease Study indicate that from 1990 to 2021, there was a marked upward trend in both the global incidence and mortality of PC, a trend expected to persist until 2050 [1, 2]. This increase is particularly pronounced in high‐income countries, attributed in part to an aging population and the rise in lifestyle‐related risk factors such as hyperglycemia and high body mass index (BMI) [3, 4]. In recent years, research on biomarkers has provided new hope for the early detection and treatment of PC. Early detection is critical, as the disease is often diagnosed at an advanced stage, resulting in poor treatment outcomes. A study on a serum proteomic biomarker panel, developed through diagnostic‐specific integrated learning, demonstrated superior predictive ability compared to the traditional cancer antigen 19–9 (CA19‐9) biomarker, underscoring the significance of novel biomarkers in early detection [5]. In terms of treatment, biomarkers are equally crucial. The Aurora A‐TPX2 interaction inhibitor has shown efficacy as a targeted anti‐mitotic agent in vivo, suggesting its potential application in treating PC [6]. These findings suggest that integrating new biomarkers into clinical practice can facilitate more personalized treatment decisions, thereby improving treatment outcomes. The integration of Mendelian randomization (MR) analysis with metabolomics has emerged as a pivotal direction in contemporary biomedical research, particularly for identifying tumor biomarkers. This method enables researchers to pinpoint disease‐related metabolites and genes. For instance, a study on ovarian cancer employed MR analysis to uncover a causal link between gut microbiota, metabolites, and ovarian cancer risk, offering a novel perspective for early detection and diagnosis [7]. Similarly, this analysis has identified protein biomarkers linked to breast cancer, providing fresh insights into its diagnosis and treatment [8]. As a tool for causal inference, MR analysis employs genetic variations as instrumental variables to estimate the causal connection between exposure factors and diseases, thus offering robust evidence for the discovery of tumor biomarkers. The role of phenylalanine in tumors is a complex and multifaceted subject, encompassing its involvement in metabolic pathways and its potential influence on tumorigenesis and progression. As an aromatic amino acid, phenylalanine serves not only as a fundamental building block for protein synthesis but also as a key participant in numerous critical biochemical reactions and metabolic pathways. Its significance in tumor metabolism is evident in its role within aromatic amino acid metabolic pathways. For instance, a study on allergic reactions to liposomal adriamycin in breast cancer patients revealed disturbances in phenylalanine metabolism, alongside other aromatic amino acids like tyrosine and β‐alanine. This finding suggests phenylalanine's crucial role in tumor metabolic networks [9]. Moreover, phenylalanine's involvement in tumors can be understood through its impact on the immune response. Research indicates that phenylalanine metabolites may affect prostate cancer risk by modulating inflammatory marker expression, highlighting its potential role in the tumor immune microenvironment (TME) [10]. Additionally, phenylalanine's influence extends to drug response in tumors, further underscoring its multifaceted role [11]. The application of machine learning (ML) algorithms in tumor research is increasingly prevalent, significantly enhancing the diagnosis, prognosis prediction, and personalized treatment of tumors. By integrating diverse data sources and analytical methods, ML offers more precise tools for tumor classification and prognosis assessment. For example, ProteoBoostR utilized the XGBoost algorithm to classify proteomic data from glioblastoma and lung adenocarcinoma, showcasing the potential of ML in integrating multi‐omics data [12]. Additionally, in gastric cancer, ML models based on the logarithmic probability of positive lymph nodes have markedly improved survival prediction accuracy, highlighting the benefits of ML in analyzing complex clinical data [13]. Building on previous research, the integration of metabolomics, MR, and ML was conducted to screen candidate biomarkers, significantly improving the biological interpretability and robustness of results. Our aim is to determine differentially expressed genes (DEGs) related to causally related differential metabolites and create ML models, thus offering new strategies for the early detection of PC and elucidating its biological mechanisms. We conducted a quantitative analysis of 55 plasma metabolites to identify differential metabolites associated with PC and confirmed their causal relationships using MR. Phenylalanine emerged as a key metabolite. We explored its regulatory genes using the GeneCards database and cross‐referenced them with PC differential genes from the Gene Expression Omnibus (GEO). This process yielded eight PC differential genes linked to metabolites. We then developed a machine‐learning diagnostic model based on these genes, integrating it with immune infiltration scores to enable precise stratified diagnosis of PC. This approach offers a novel method for more accurate gene‐level diagnosis of PC. Additionally, we created a scoring system to quantify metabolite‐related PC differential gene patterns in patients with varying risk levels, predicting their sensitivity to drug treatments. This study demonstrates that phenylalanine‐related genes are pivotal in PC, aiding in prognosis prediction and providing personalized treatment options for individuals with PC. Figure 1 presents the workflow. Research design flowchart. Materials and Methods Public Sample Collection and Processing We systematically gathered publicly available, clinically annotated datasets on PC gene expression. Patients lacking survival information were excluded from further analysis. This study included six sample cohorts of PC patients: GSE16515, GSE62452, GSE183795, GSE15471, GSE28735, GSE71989. GSE16515, GSE62452, and GSE183795 constituted the training set for subsequent model establishment. GSE15471, GSE28735, and GSE71989 served as the validation set. We downloaded microarray gene expression datasets from the Affymetrix and Illumina platforms via the GEO database (https://www.ncbi.nlm.nih.gov/geo/). To determine DEGs between normal and PC samples, we first merged the datasets. Then, principal component analysis (PCA) was conducted on the original data to decrease dimensionality and observe sample distribution in high‐dimensional space. Utilizing the merged dataset, we performed differential expression analysis with “limma” to identify DEGs and convert probe IDs into gene symbols. The p‐value was adjusted using the FDR (Benjamini‐Hochberg) method, and genes were screened depending on FDR < 0.05 and |log2FC| ≥ 1. Lastly, the “ggplot2” and “pheatmap” packages were employed to generate volcano plots and heat maps for visualization. Metabolomics Differential Analysis This investigation utilized data from the Metabolomics Workbench, part of the National Metabolomics Data Repository (NMDR) under the National Institutes of Health (NIH) in the United States. We selected various differential metabolic characteristics across distinct metabolic subtypes to identify biomarkers for PC. The dataset, identified by project number PR001339, is accessible via the project DOI: 10.21228/M8N405. Unsupervised PCA visualizes intrinsic structures, temporal trajectories, and potential outliers. To particularly examine supplementation‐correlated divergence throughout recovery, we utilized orthogonal partial least squares discriminant analysis (OPLS‐DA) by the ropls package. The counts of prognostic and orthogonal constituents were detected through cross‐validation (CV) within the ropls framework. We evaluated model performance via R 2 Y and Q 2 metrics with 7‐fold CV, while model stability was estimated via permutation testing (1000 permutations). Discriminant metabolites were determined through a combined multivariate‐univariate framework: features with Variable Importance in Projection (VIP) scores greater than 1.0 were prioritized and subsequently filtered using false discovery rate (FDR) < 0.05 to control for false discoveries [14]. The optimal model retained one predictive component and two orthogonal components, as determined by 7‐fold cross‐validation within the ropls framework. GWAS Data for PC GWAS data for PC were acquired from OpenGWAS. MR analyses of plasma metabolites and PC used dataset ieu‐a‐822, which comprised 3835 participants (1896 cases and 1939 controls). Phenylalanine was treated as the exposure variable and PC as the outcome. The study followed the three core assumptions of MR: (i) instrumental variables are strongly related to the exposure; (ii) instrumental variables are independent of confounders; and (iii) instrumental variables influence the outcome only through the exposure. SNPs were selected as instrumental variables using a significance threshold of p < 1 × 10−5. Linkage disequilibrium pruning was performed using a 100‐kb window and r 2 < 0.1 to minimize population stratification effects. SNPs associated with known confounders or outcomes were excluded based on phenotype database information. Instrument validity was further assessed using the MR‐PRESSO outlier test and MR‐Egger intercept test (α = 0.05). For causal inference, the IVW technique was utilized as the primary analysis, supplemented by weighted median, MR‐Egger regression, and simple and weighted modes methods. Cochran's Q test assessed heterogeneity, while leave‐one‐out sensitivity analysis evaluated result robustness. Potential horizontal pleiotropy was assessed via both the MR‐Egger intercept and the MR‐PRESSO global tests. Metabolite‐Related Gene Data Acquisition As a comprehensive gene database, GeneCards integrates data from about 150 sources to provide detailed annotations for genes, including function, disease associations, expression levels, and pathways (https://www.genecards.org/) [15]. It is an important tool in biomedical investigation. We utilized the GeneCards database to obtain genes related to phenylalanine. Weighted Gene co‐Expression Network Analysis We created a weighted gene co‐expression network for DEGs using the R “WGCNA” (v1.71) to screen functional gene modules highly related to PC [16]. A soft‐thresholding power (β) was evaluated from a range of 1 to 20. A power of β = 20 was selected as the lowest value at which the scale‐free topology fit index (R 2) exceeded 0.85, ensuring an approximate scale‐free network topology. First, determine the soft threshold power (β = 1 to 20) to create a scale‐free topological structure. Subsequently, a weighted adjacency matrix is produced and converted into a topological overlap matrix (TOM). In addition, dissTOM is acquired through hierarchical clustering, and the dynamic tree‐cutting technique is adopted to screen each module based on gene‐similarity clustering. Herein, the “minModuleSize” and “MEDissThres” are set to 60 and 0.25, respectively. Subsequently, we associated the identified modules with two traits (the normal and the PC group). Select the genes highly correlated with PC in the module for further analysis. We conducted subsequent studies by taking the intersection of genes related to phenylalanine metabolism and DEGs of PC through Venn diagrams. The Venn diagram was constructed via the VennDiagram [v1.7.3] package. ML Model Development and Interpretation ML integration framework incorporated 12 algorithms: Least Absolute Shrinkage and Selection Operator (Lasso), Ridge Regression (Ridge), Elastic Net (Enet), stepwise generalized linear model (Stepglm), glmBoost, linear discriminant analysis (LDA), partial least squares regression for generalized linear models (plsRglm), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and Naive Bayes. A total of 113 algorithm combinations were generated using 10‐fold CV, with one algorithm applied for selecting features and another for predictive model construction. Expression profiles from the training and testing cohorts were first standardized using Z‐score transformation. Four feature‐selection algorithms (Lasso, Stepglm, RF, and glmBoost) were then used to identify candidate genes, followed by the construction of prediction models using Ridge, Enet, LDA, plsRglm, GBM, XGBoost, SVM, and Naive Bayes. Hyperparameter optimization and model fitting were performed within the 10‐fold CV framework. Model performance was estimated in both cohorts using the area under the receiver operating characteristic (ROC) curve (AUC), and the model with the highest average AUC across the training and test sets was selected as the optimal model. Characteristic genes were then identified from this model. Confusion matrices were generated to assess sensitivity, specificity, and overall accuracy. Correlations among hub genes were analyzed and observed via the R “PerformanceAnalytics” (v2.0.4) [17]. To improve model interpretability and address the “black box” nature of ML methods, SHAP (Shapley Additive exPlanations) analysis was used to quantify each feature's contribution to model predictions. SHAP values were computed using the TreeSHAP algorithm, as the optimal model was a tree‐based random forest. The background dataset was constructed by randomly sampling 100 instances from the training set to serve as the reference distribution for SHAP value computation. This method was chosen to improve computational efficiency and reduce variance, especially in our high‐dimensional dataset. Nomogram Development and Verification We conducted Cox regression analyses (univariate and multivariate) on clinical variables (e.g., risk score, age, stage, grade, and BMI), evaluating the prognostic significance of each variable by hazard ratios and p‐values. A nomogram was constructed using the “rms” and “regplot” packages. The nomogram's prognostic accuracy for overall survival in PC patients was assessed with calibration curves. Calibration of the nomogram was further examined by calibration plots and the Hosmer–Lemeshow (H–L) test (p > 0.05, good calibration). Decision curve analysis (DCA) was employed to estimate the net benefit of the model across a range of threshold probabilities. Gene Set Enrichment Analysis (GSEA) and Gene Set Variation Analysis (GSVA) We assessed the model construct genes through GSEA [18]. The gene set for the Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathways was sourced from the Molecular Signatures Database, specifically the Human MSigDB database (c2.cp.KEGG. Hs.symbols.gmt). Pathways were considered significant when they had normalized enrichment scores (NES) > 1 and p < 0.05. Additionally, we calculated final scores using GSVA v1.47.0 to facilitate further analysis. Immune Infiltration The study assessed the relative abundance of 22 immune cell subtypes in both normal and PC samples using the CIBERSORT algorithm. Following this, the relationship between PC construct model genes and immune cells was evaluated through Spearman correlation analysis. To detect the relationship between immune cell subtypes and the predicted quantities of key genes, the “corrplot” R package was employed, utilizing the Spearman rank correlation method. The findings were visualized using the “ggplot2” and “ggpubr” R packages [19]. To evaluate the immunological significance of the key genes, we integrated their expression data with immune cell infiltration outcomes. Spearman correlation analysis quantified the association between gene expression and immune cell abundance. A coupled correlation network between immune cells and key genes was constructed using linkET. This network was visualized through coupled correlation graphs, where the thickness and color of the edges represented the robustness and direction of these relationships. Analysis of Single‐Cell Sequencing Data Single‐cell RNA‐sequencing datasets for PC (GSE197177) were obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/, accessed on 23 August 2023) [20]. The GSE197177 dataset, generated on the GPL18573 platform, was integral to our analysis. This dataset included one adjacent normal tissue sample as the control group and three primary untreated pancreatic ductal adenocarcinoma tumor samples as the experimental group. Subsequently, differential gene expression analysis was performed on the single‐cell sequencing data using R. The analysis employed parameter settings of |log2‐fold change| ≥ 1 and adjusted p < 0.05. Single‐Cell Simulation of Gene Knockout (scTenifoldKnk) and Gene Enrichment Analysis We utilized the scTenifoldKnk computational tool [21] to perform virtual knockout (KO) analysis, simulating regulatory perturbation effects in single‐cell RNA sequencing data. Using scTenifoldKnk, we constructed both the baseline gene regulatory network (GRN) and the perturbed GRN for the SLC6A14 knockout state. We identified genes with significant perturbations between these networks, applying a false discovery rate (FDR) threshold of less than 0.05. This allowed us to determine the genes affected by these differences and subsequently conduct pathway enrichment analysis. Additionally, we performed Gene Ontology (GO) and KEGG analyses on the entire set of ranked genes to assess the broader functional changes associated with SLC6A14 perturbations in PC. Survival Curves, Prognostic Analysis, and ROC Curves Associated With the SLC6A14 Gene An analysis of pancreatic cancer (PC) clinical data from the TCGA database revealed an association between the SLC6A14 gene and survival curves for PC. Additionally, software packages including “timeROC,” “survminer,” “survival,” and “reshape2” facilitated univariate and multivariate prognostic analyses, as well as survival analysis. Drug Prediction We used the Enrichr platform DSigDB (http://dsigdb.tanlab.org/DSigDBv1.0/) for drug‐gene interaction analysis [22]. To determine candidate drugs targeting key genes (COL17A1, GJB2, MET, SLC2A1, and SLC6A14). Structural information on bioactive compounds is derived from the database of active compounds in traditional Chinese medicine (http://hit2.badd‐cao.net/). Molecular Docking and Molecular Dynamics Simulation We used GROMACS 5.0.4 to perform molecular dynamics simulations on the receptor protein with the highest ligand‐binding energy from the molecular docking results. The CHARMM36m force field was applied to model the protein, and the genistein ligand was parameterized using the CGenFF server. The protein–ligand complex was solvated in a cubic box with the TIP3P explicit water model, and Na+/Cl− counterions were added to neutralize the system. Energy minimization was first performed using the steepest descent algorithm. The system was then equilibrated under an NVT ensemble for 100 ps, followed by an NPT ensemble for 100 ps, with position restraints applied to the protein backbone. Temperature was maintained at 310 K using the V‐rescale thermostat, and pressure was controlled at 1 bar using the Parrinello–Rahman barostat. Electrostatic interactions were computed using the Particle Mesh Ewald (PME) method with a cutoff of 1.4 nm, and the same cutoff was applied for van der Waals interactions. All covalent bonds involving hydrogen atoms were constrained using the LINCS algorithm, allowing an integration time step of 2 fs. Following equilibration, a 100‐ns production MD simulation was performed under the NPT ensemble without restraints, and trajectory frames were saved every 10 ps for analysis. Cell Culture, RNA Isolation and Quantitative Real‐Time PCR (RT‐qPCR) The PC cell line (BxPC‐3) and the normal pancreatic epithelial cell line HPDE6‐C7 were bought from Henan Provincial Industrial Microbial Strain Engineering and Technology Research Center (China). BxPC‐3 was cultivated in RPMI‐1640 medium (HyClone, Logan, UT, USA) enriched with 10% fetal bovine serum (Gibco, USA), and HPDE6‐C7 was cultivated in DMEM medium (HyClone, Logan, UT, USA) enriched with 10% fetal bovine serum (Gibco, USA). The incubation of all cells was conducted at 37°C with 5% CO2. RT‐qPCR operation was conducted as previously described. Total RNA was extracted from cells using TRIzol reagent (Vazyme, China, R711‐01). 1 μg total RNA was reverse‐transcribed into cDNA using a reverse transcription kit (Vazyme, China, R333‐01). We conducted RT‐qPCR via the ChamQ SYBR Color qPCR Master Mix (Vazyme, China, Q411‐02). Ultimately, we standardized the gene levels to β‐actin and assessed them via the 2−ΔΔCT technique. Data S1 lists the primers used. Results GEO Screening of DEGs To determine DEGs, we combined samples from the three datasets. We then conducted PCA on the raw data to decrease dimensionality and observe sample distribution in high‐dimensional space. The PCA plot revealed that sample points were relatively well mixed across datasets and were no longer segregated by source (Figure 2A). Utilizing the “limma” with selection criteria of a FDR, we identified 309 DEGs. Among these, 191 genes were overexpressed, and 118 genes were downexpressed (Figure 2B). Differentially expressed gene Analysis. (A) PCA plot after batch processing; (B) Volcano plot of DEGs. Development of the co‐Expression Network and Determination of Core Genes We conducted WGCNA on the merged dataset to create the gene co‐expression network. A soft‐thresholding power of 20 was selected to meet the scale‐free topology criterion, resulting in an R^2 of 0.9 (Figure 3A). This analysis identified nine modules (Figure 3B). We visualized the module classifications and gene co‐expression patterns using a dendrogram and a heatmap (Figure 3C). Furthermore, module–trait correlation analysis illustrated a strong positive link between the red module and PC with an r value of 0.72 and a p‐value of 1e‐80 (Figure 3D). Consequently, we concentrated subsequent analyses on the red module, as it more specifically indicates PC and merits further detailed investigation. In addition to its strong statistical correlation, functional enrichment analysis of the 498 genes in the red module revealed significant over‐representation of pathways closely associated with pancreatic cancer pathogenesis. GO enrichment analysis revealed significant enrichment of these genes in biological processes such as positive regulation of cell cycle, cell‐matrix adhesion, and cadherin binding. Moreover, KEGG pathway analysis showed significant enrichment in pathways including ECM–receptor interaction, focal adhesion, PI3K–Akt signaling, HIF‐1 signaling, and glycolysis/gluconeogenesis (Data S2). These pathways are intimately linked to key hallmarks of PC, such as stromal remodeling, metabolic reprogramming, and hypoxia adaptation, further supporting the biological relevance of the red module as a PC‐associated gene set. Construction of WGCNA. (A) Soft threshold power (left) and average connectivity (right) of WGCNA; (B) Heatmap of correlation between module eigengene (ME) and clinical phenotype; (C) Gene dendrogram with multiple partitioning modules; different colors represent different clusters (modules); (D) Scatter plot of correlation between module membership (MM) and gene significance (GS) of red module. Selection and Functional Enrichment Analysis of Phenylalanine‐Related Genes in PC To examine the regulatory function of phenylalanine in the pathogenesis of PC, we first conducted an intersection analysis among DEGs, genes in the red module, and phenylalanine‐related genes, yielding eight signature genes (Figure 4A). We then conducted functional enrichment analyses on these signature genes to explore phenylalanine‐associated biological functions and potential pathways in PC. GO enrichment analysis indicated that the signature genes primarily influence chemotaxis (positive chemotaxis, cellular response to chemical stress), transmembrane transport (carboxylic acid transmembrane transport, organic acid transmembrane transport), and aromatic amino acid transmembrane transporter activity (Figure 4B). KEGG pathway analysis revealed the highest enrichment of these signature genes in cancer pathways (e.g., Renal cell carcinoma, Central carbon metabolism in cancer, Melanoma, Non − small cell lung cancer); drug resistance pathways (Platinum drug resistance), and indirectly illustrates that phenylalanine may be a potential key component in the pro‐tumorigenic mechanisms during PC. Functional enrichment analysis. (A) Venn diagram of candidate genes. (B) GO functional enrichment analysis of candidate genes. (C) KEGG pathway enrichment analysis of candidate genes. Metabolomics Analysis of Differential Metabolites To identify differential lipid metabolites, metabolomic profiling was conducted on the dataset. Figure 5A illustrates that the PCA results revealed a separation between the control and disease groups along PC1 (principal component). This pronounced difference in metabolic profiles suggests that patients' metabolic features alter with the progression of PC. To further highlight intergroup differences, we utilized OPLS‐DA and generated a score plot (Figure 5B). OPLS‐DA combines an orthogonal signal correction (OSC) filter with PLS‐DA, decomposing the X matrix into Y‐related and Y‐unrelated information, thereby eliminating variation unrelated to classification. We evaluated the model's predictive ability and statistical significance using 1000 permutation tests. The OPLS‐DA model produced a Q2 value of 0.954 from CV (Data S3), confirming the model's suitability. The OPLS‐DA score plot clearly demonstrated a division between the two comparison groups. Differential metabolites in metabolomics. (A) Principal component analysis (PCA) plot demonstrating distinct clustering of sample groups. (B) OPLS‐DA scores plot for discriminating between PC and Normal groups. (C) Heatmap of differential metabolites between Normal and PC groups. (D) Integrated visualization of pathway analysis for key differential metabolites. The volcano plot illustrates that, using the criteria of VIP > 1.0, FDR < 0.05, and logFC > 0.585, 12 differential lipid metabolites were identified between the PC group and controls. Among these metabolites, three were downregulated, while nine were upregulated in the PC group. A heatmap using color gradients depicted the variation in measurements of these 12 differentially expressed metabolites between the PC and control groups (Figure 5C). KEGG enrichment analysis illustrated that these lipid metabolites were significantly enriched in several biological processes and signaling pathways, including ABC transporters, choline metabolism in cancer, and central carbon metabolism in cancer (Figure 5D). In conclusion, we determined 12 differentially expressed metabolites and performed preliminary investigations into their biological functions. MR Probes Differential Metabolites With Causal Connections Through integrated metabolomic profiling and Mendelian randomization (MR), phenylalanine was identified as a differentially abundant plasma metabolite with a causal association with PC. Using 15 genome‐wide significant SNPs (p < 1 × 10−5) as instrumental variables, the inverse‐variance weighted (IVW) method—the primary analysis—confirmed a significant causal effect (OR = 1.641, 95% CI: 1.052–2.562, p = 0.029), with supportive results from MR‐Egger, weighted median, simple mode, and weighted mode (Figure 6A,B,G). Sensitivity analyses further validated these findings. Cochran's Q test revealed no significant heterogeneity among SNPs (IVW: Q = 6.764, df = 14, p = 0.943; MR‐Egger: Q = 6.034, df = 13, p = 0.945; both p > 0.05). The MR‐Egger intercept test detected no evidence of directional horizontal pleiotropy (intercept = −0.044, SE = 0.051, p = 0.408). The funnel plot exhibited a symmetric distribution of individual SNP estimates (Figure 6C), and leave‐one‐out analysis confirmed that no single SNP disproportionately influenced the overall causal estimate (Figure 6D). ROC analysis demonstrated excellent diagnostic performance of phenylalanine for PC (AUC = 1.000; Figure 6E), and boxplots confirmed significantly elevated phenylalanine levels in PC versus normal tissues (Figure 6F). Together, these results establish phenylalanine as a robust, causally implicated metabolite in PC pathogenesis. MR screening for potential metabolic markers of PC. (A–D) Scatter plots, forest plots, and funnel plots for phenylalanine, showing the causal effects on PC risk, evaluating the causal effects of individual SNPs, and assessing overall heterogeneity. Leave‐one‐out analysis further visualizes the independent impact of each gene on PC risk. (E) Diagnostic performance of phenylalanine. (F) Expression levels of phenylalanine. (G) Forest plot showing p‐values from IVW analysis for the screened phenylalanine. ML Diagnostic Model Development To develop an effective PC diagnostic model, we utilized several ML algorithms and their stochastic combinations. We combined data from the training and test cohorts, then split the dataset using a random seed, allocating 70% for training and 30% for validation. An independent validation cohort assessed the model's generalizability. We averaged and ranked AUC values across all cohorts, using the mean AUC as a complex performance metric to estimate model stability and generalization. The model with the highest mean AUC, RF, was selected as optimal (Figure 7A). In the training set, this model reached an AUC of 0.994, while in the validating sets, AUCs were 0.918, 0.983, and 0.923 (Figure 7B). Confusion matrices for the five genes demonstrated favorable diagnostic value (Figure 7C). In summary, we established an RF diagnostic ML model, and these five genes showed strong potential for distinguishing between normal and PC states. Model Construction. (A) ROC curves for 113 algorithms in machine learning. (B) ROC curve of the optimal model. (C) Confusion matrices for each dataset under the best model. Verification of Diagnostic Effectiveness by Logistic Regression ML To comprehensively evaluate the diagnostic performance of the five selected genes, normal and PC samples were analyzed using a logistic regression model. DCA illustrated that the five‐gene model presented a higher net benefit across most threshold probabilities under varying cost–benefit conditions (Figure 8A). A nomogram was constructed depending on the five genes to predict PC risk, illustrating the contribution of each predictor (including SLC6A14 and the remaining genes) to the total risk score (Figure 8B). Calibration curves further confirmed good agreement between expected and detected outcomes (Figure 8C). Overall, the model showed strong diagnostic performance and may support early detection and intervention of PC through quantitative assessment of gene expression levels. The results of Logistic machine learning. (A) Decision Curve Analysis (DCA), the red, gray, and black lines represent the net benefit of the 5‐gene model, the scenario where all samples receive treatment, and the scenario where no samples receive treatment, respectively. (B) Results of nomogram are used to predict the risk of PC. (C) The calibration curves for the radiomic nomogram in the primary and validation cohorts. (D) Pearson correlation analysis on five genes expression profiles. (E) The box plot shows the expression of five genes in the normal group and the PC group. (F) ROC curves for five genes. ***: p < 0.001. To further elucidate the cooperative mechanisms by which the five model‐selected genes contribute to the initiation and progression of PC, we first performed Pearson correlation analysis on their expression profiles (Figure 8D). The outcomes illustrated a highly significant co‐expression pattern among these core genes. Concurrently, these genes were found to be highly expressed in PC (Figure 8E). This tightly regulated network suggests that they may jointly participate in key biological pathways driving disease progression and function cooperatively as a potential gene module. Based on the high concordance exhibited by these genes at the molecular level, we further evaluated their potential utility as individual or combined biomarkers for clinical diagnosis. By constructing ROC curves, we systematically assessed the ability of these genes to discriminate disease samples from normal controls (or high−/low‐risk subgroups) (Figure 8F). In the ROC space, with 1 − Specificity (false‐positive rate) on the x‐axis and Sensitivity (true‐positive rate) on the y‐axis, the core gene model demonstrated excellent diagnostic performance. These results indicate that these genes have substantial translational potential as novel noninvasive clinical diagnostic markers. Feature Explanation of the Model Depending on SHAP Using the trained RF model, SHAP analysis was performed to evaluate feature importance. The SHAP summary plot showed that SLC6A14, SLC2A1, and GJB2 were the top three genes with the greatest contributions across models (Figure 9A). The SHAP swarm plot further illustrated the distribution and direction of the contribution of all genes to model predictions, indicating that higher levels of these three genes were related to an elevated incidence of PC (Figure 9B). Figure 9C illustrates that the features “SLC6A14 = 2.53”, “GJB2 = 3.63”, “SLC2A1 = 5.54”, “MET = 5.47”, and “COL17A1 = 4.49” contribute negatively to the model's prediction of PC, with SHAP values of −0.169, −0.144, −0.14, −0.0743, and −0.0505, respectively. This indicates that lower values of these features increase the probability that a sample is predicted as PC. These findings provide a perspective on feature effects that goes beyond mere importance rankings. SHAP analysis. (A, B) Contribution distribution and direction of 5 genes. (C) Probability map of 5 genes predicting PC. Immune Filtration Analysis To explore the possible links between immune alterations in PC and core DEGs, we utilized CIBERSORT to assess immune cell distributions in PC and normal control samples. We visualized the results as percentage bar plots (Figure 10A). Compared to controls, the PC patient cohort showed significant increases in Macrophages (M0, M1, and M2) and resting Dendritic cells (resting and activated), while CD8+ T cells, Monocytes, and Naïve B cells significantly decreased (Figure 10B). Additionally, a significant positive correlation emerged between Activated Mast cells and γδ T cells in PC. Conversely, Resting Memory CD4+ T cells illustrated a negative link with Activated NK cells (Figure 10C). A gene–immune cell association network (Figure 10D) further revealed that Macrophages M0, resting NK cells, and activated Dendritic cells are significantly associated with multiple genes. This suggests that these genes may play roles in immune regulation and polarization, antigen presentation, initiation of adaptive immunity, and innate immune surveillance and early warning. CD8+ T cells exhibited significant negative correlations with most genes, indicating their expression might be linked to tumor cell clearance. Analysis of immune cell infiltration and inflammatory five genes in PC. (A) Relative distribution of 22 immune cell types in normal and PC samples. (B) Differential analysis of 22 immune cell types in normal and PC samples. (C) Correlation heatmap of immune cell subpopulations. (D) Analysis of the relationships between five genes and immune cells. Single‐Cell Sequencing Analysis Revealed the Distribution of SLC6A14 in Immune Cells We obtained single‐cell sequencing data for pancreatic ductal adenocarcinoma from the GEO database (GSE197177) for our analysis. Initially, we performed several routine analyses, including quality control, sequencing depth, differential gene, and principal component analyses (Data S4). Cluster analysis revealed that all cells were grouped into 18 distinct populations, such as B cells, CD8+ T cells, macrophages, endothelial cells, monocytes, fibroblasts, and epithelial cells (Figure 11A,B). The principal component analysis for both tumor and normal groups displayed similar cell clusters (Figure 11C). Following this, we conducted a cell communication analysis to explore signal transduction mechanisms among different cell types. In the resulting figure, each node represents a cell type, with larger nodes indicating a greater number of cells. The analysis demonstrated that interactions among adipocytes, epithelial cells, fibroblasts, and other cells were most frequent (Figure 11D), with the strongest interaction intensity observed among these cell types (Figure 11E). Correlation analysis further showed that the SLC6A14 gene was predominantly expressed in epithelial cells (Figure 11F). These findings suggest that in PC tissues, SLC6A14 is closely associated with epithelial cells. Notably, epithelial cells—the primary cell type expressing SLC6A14 (Figure 11F)—exhibited extensive intercellular communication with multiple immune cell populations, including macrophages and CD8+ T cells. Given that SLC6A14 functions as an amino acid transporter capable of mediating phenylalanine transmembrane transport [38], its predominant expression in epithelial cells within the tumor microenvironment suggests a potential role in reprogramming amino acid metabolism and modulating immune cell function through altered nutrient availability. The graphical representation of single‐cell sequencing data analysis results includes several key diagrams: (A) a cluster analysis diagram, (B) a cell annotation diagram, and (C) a principal component analysis diagram, which also features a comparison between the tumor and control groups. Additionally, (D) an interaction quantity graph illustrates the number of cells involved in interactions, with thicker connections indicating greater involvement. (E) An interaction intensity diagram shows the strength of these interactions, where thicker lines denote higher intensity. Finally, (F) a correlation analysis graph presents the relationship between SLC6A14 and various immune cells. Simulated SLC6A14 Knockout Revealed Its Underlying Molecular Mechanism To delve deeper into the molecular mechanisms of the SLC6A14 gene in pancreatic cancer pathology, we performed a single‐cell omics gene knockout simulation analysis. Our findings revealed that knocking out the SLC6A14 gene significantly impacted 95 genes, as detailed in Figure S2. Notably, the most affected genes included TYROBP, FCER1G, ITGB2, KLK1, and FXYD3 (Figure 12A,B). These significantly impacted genes represented 1.9% of the total gene pool (Figure 12C). Impact of SLC6A14 Gene Knockout on Gene Expression. (A) Bar chart illustrating genes significantly impacted by the knockout of SLC6A14. (B) Volcano plot depicting the genes significantly influenced by the SLC6A14 knockout. (C) Pie chart showing the distribution of genes significantly affected by the SLC6A14 knockout. (D) Enrichment analysis of genes significantly altered following the simulated knockout of the SLC6A14 gene. Following this, we conducted GO and KEGG enrichment analyses on the identified genes. The GO enrichment analysis demonstrated a strong correlation between the 95 genes and processes such as cell–cell junction organization, neutrophil activation, tight junction formation, and cadherin binding. Meanwhile, the KEGG analysis highlighted pathway enrichments in tight junctions, pathogenic Escherichia coli infection, cell adhesion molecule (CAM) interactions, and leukocyte transendothelial migration (Figure 12D). These pathways are closely linked to cell migration, bacterial infection, and immune activation, suggesting that the knockout of the SLC6A14 gene influences these tumor‐related pathways. Determination of Candidate Drugs, Molecular Docking, and Molecular Dynamics Simulation From the DSigDB database on the Enrichr platform, 110 drug molecules were found and compared with a database of active compounds from traditional Chinese medicine, 19 candidate herbal extracts (Figure 13A). Since SLC6A14 had the highest model‐predicted score, we performed molecular docking on it. The complete docking results for all 19 candidate compounds are provided in Data S5. Genistein and withaferin A had higher affinities for SLC6A14 of 9.1 and 8.9 kcal/mol, respectively; here we present the docking result for genistein with SLC6A14 (Figure 13B). We performed a 100‐ns molecular dynamics simulation of genistein bound to SLC6A14 to evaluate the stability of interactions and the binding affinity of genistein as a bioactive regulator of SLC6A14 protein. The docking and molecular dynamics simulation of SLC6A14‐genistein. (A) Prediction of candidate drugs. (B) Molecular docking of SLC6A14 with genistein. (C) Rg values of the SLC6A14 with genistein complex over time. (D) Number of HBonds in the SLC6A14 with genistein complex over time. (E) RMSD values of the SLC6A14 with genistein complex over time. (F) RMSF values of the SLC6A14 with genistein complex. (G) SASA values of the SLC6A14 with genistein complex over time. (H) Analysis of protein‐ligand MMPBSA (KJ/mol). Analysis of the radius of gyration (Rg) illustrated that the SLC6A14 protein remained highly compact, measuring 2.80 ± 0.03 nm throughout the 100 ns simulation, with no significant structural changes observed (Figure 13C). Hydrogen bond analysis showed that, on average, 1–2 hydrogen bonds (1.45 ± 0.99) formed between the protein and genistein, ensuring stable non‐covalent interactions (Figure 13D). The average root mean square deviation (RMSD) for the protein‐small molecule complex was 0.91 ± 0.10 nm, while the protein alone had an RMSD of 0.86 ± 0.10 nm, and the Genistein drug exhibited an RMSD of only 0.05 ± 0.01 nm. These low and stable RMSD values indicate that the complex system quickly reached equilibrium at the start of the simulation and kept a stable conformation for the subsequent 100 ns (Figure 13E). The average root‐mean‐square fluctuation (RMSF) was 0.18 ± 0.17 nm, reflecting the good rigidity of the protein residues and limited dynamics in the binding region (Figure 13F). The solvent‐accessible surface area (SASA) remained stable at 286.91 ± 6.84 nm2, suggesting no significant changes in the protein's surface properties (Figure 13G). The binding free energy, calculated using MM/GBSA, was −33.77 ± 1.08 kcal/mol, with van der Waals (−36.62 ± 0.76 kcal/mol) and electrostatic interactions (−38.54 ± 0.11 kcal/mol) being the primary contributors. This indicates a strong binding affinity between genistein and SLC6A14 (Figure 13H). These results support genistein's role as a modulator of SLC6A14, presenting a basis for further research and drug development. The SLC6A14 Gene Have an Impact on the Survival of PC Patients After analyzing the cell distribution and influence pathways in single‐cell sequencing data containing the SLC6A14 gene, we evaluated its impact on patient survival prognosis. The results indicated that the SLC6A14 gene was significantly upregulated in PC tumor tissues, correlating with lower patient survival rates (Figure 14A,B). Additionally, the ROC curve, with an AUC of 0.961, demonstrated the gene's strong diagnostic potential (Figure 14C,D). RT‐qPCR test results further confirmed that all five genes (SLC6A14, SLC2A1, MET, GJB2, and COL17A1) involved in the modeling were highly expressed in PC cells (Figure 14E). The SLC6A14 gene can serve as an impact on PC. (A) The expression level map of SLC6A14 in tumor tissues and normal tissues in TCGA data. (B) Overall survival analysis graph (Kaplan–Meier curve) for the high and low expression groups of the SLC6A14 gene. (C) The ROC curve for predicting the overall survival rate of PC patients based on the expression level of SLC6A14. (D) The time‐dependent ROC curve for predicting overall survival in PC patients based on SLC6A14 expression levels. (E) The expression levels of SLC6A14, SLC2A1, MET, GJB2, and COL17A1 mRNA were detected using RT‐qPCR. *: p < 0.05; **: p < 0.01; ***: p < 0.001. Discussion PC is a highly invasive malignancy of the digestive system with a 5 year survival rate below 10%, largely attributable to late‐stage diagnosis [23]. Current serological markers, such as CA19‐9, and imaging modalities lack the sensitivity and specificity required for early detection [24]. Consequently, identifying novel molecular biomarkers is essential for improving early screening and patient prognosis [25]. In this study, we integrated metabolomics, Mendelian randomization (MR), and machine learning (ML) to develop a phenylalanine metabolism‐based diagnostic model for PC. Metabolomic profiling identified phenylalanine as a differentially abundant plasma metabolite, and MR analysis confirmed its causal association with PC risk. WGCNA and differential expression analysis identified eight phenylalanine‐related hub genes, among which an RF‐based ML model selected five core genes—COL17A1, GJB2, MET, SLC2A1, and SLC6A14—as the most discriminative features. The RF model demonstrated excellent diagnostic performance, achieving a mean AUC of 0.955 across 113 algorithm combinations. From the confusion matrices, the RF model achieved a sensitivity of 94.7%, specificity of 94.0%, and overall accuracy of 94.4% on the training set. Across the three independent validation cohorts, the model maintained robust performance with sensitivities of 87.2%, 95.6%, and 84.6%; specificities of 87.2%, 88.9%, and 100.0%; and accuracies of 87.2%, 92.2%, and 90.5%, respectively (Data S6). Additionally, the GSE71989 validation cohort comprised a relatively small sample size (n = 21), which may limit the generalizability of the performance metrics obtained from this dataset. A gap between training AUC (0.994) and validation AUCs (0.918–0.983) warrants consideration of potential overfitting. However, 10‐fold cross‐validation was employed during training, and the model was validated on three independent external cohorts, with consistently high performance (> 0.91 AUC across all cohorts) confirming robust generalizability. The ΔAUC of 0.011–0.076 falls within the expected range for high‐dimensional omics‐based ML models [26, 27]. SHAP analysis identified SLC6A14 as the top‐ranked feature driving model predictions. SLC6A14 (ATB0, +), a broad‐specificity amino acid transporter, is upregulated in PC and 11 other tumor types [28, 29]. It directly mediates phenylalanine transmembrane transport [30], and its silencing reduces proliferation, motility, and invasiveness of gemcitabine‐resistant PC cells [31]. Single‐cell analysis confirmed predominant SLC6A14 expression in epithelial cells, and simulated knockout revealed enrichment of immune‐related pathways including leukocyte transendothelial migration and cell adhesion molecules (Figures 11, 12). Notably, epithelial cells—the primary cell type expressing SLC6A14 (Figure 11F)—exhibited extensive intercellular communication with macrophages and CD8+ T cells. Given that SLC6A14 mediates phenylalanine transport, its elevated expression may deplete extracellular phenylalanine, creating a metabolically restrictive niche that impairs antitumor immune effector functions. This hypothesis is supported by the knockout‐enriched pathways involving leukocyte transendothelial migration and CAM interactions (Figure 12D), collectively suggesting that SLC6A14 may serve as a metabolic checkpoint linking amino acid availability to immune evasion in PC. Drug predictive analysis identified genistein and withaferin A as candidate SLC6A14‐targeting compounds, with molecular docking scores of −9.1 and −8.9 kcal/mol, respectively. Notably, kaempferol and myricetin—two plant‐derived flavonoids structurally related to genistein—also exhibited competitive binding affinities, suggesting that the isoflavone/flavonoid scaffold may represent a privileged chemotype for SLC6A14 targeting. These findings warrant further experimental validation of these candidate compounds in PC models. Molecular dynamics simulations confirmed stable genistein–SLC6A14 binding over 100 ns, with a calculated binding free energy of −33.77 ± 1.08 kcal/mol (MM/GBSA), suggesting genistein as a potential lead compound for therapeutic targeting of SLC6A14 in PC. Taken together, these findings delineate a coherent mechanistic axis linking genistein, SLC6A14, and phenylalanine metabolism in PC. Mendelian randomization established phenylalanine as a causal risk factor for PC (Figure 6A–D). Molecular docking and MD simulations demonstrated high‐affinity binding of genistein to SLC6A14 (docking score: −9.1 kcal/mol; ΔGainh = −33.77 kcal/mol). Based on this evidence, we hypothesize that genistein may inhibit SLC6A14‐mediated phenylalanine transport, thereby limiting phenylalanine availability for tumor metabolic reprogramming and potentially restoring antitumor immune function within the tumor microenvironment. This proposed mechanism, while supported by multi‐layered in silico evidence, requires experimental validation through functional transport assays, targeted metabolomics, and in vivo efficacy studies. Our findings align with and extend recent MR‐based studies linking circulating metabolites to PC risk [32, 33] and ML‐based diagnostic models for PC [34, 35]. However, most existing studies address metabolite prioritization and model construction as separate workflows. To our knowledge, this study is the first to integrate MR‐guided causal metabolite selection with multi‐cohort ML diagnostic modeling to investigate the phenylalanine metabolism–PC axis. This strategy reduces false‐positive metabolite associations through genetic evidence and directly links metabolic causality to clinically actionable biomarkers. Several limitations should be acknowledged. The training and validation cohorts, though derived from multiple independent datasets, have moderate sample sizes; larger prospective cohorts are needed. Functional validation of SLC6A14‐mediated phenylalanine metabolic reprogramming remains lacking—in vitro knockdown/overexpression experiments combined with targeted metabolomics are required. Molecular docking and MD simulations, while providing strong in silico evidence, require experimental confirmation of genistein–SLC6A14 binding. RT‐qPCR validation was conducted in a single cell line pair (B × PC‐3 vs. HPDE6‐C7) and should be extended. Future work will focus on: (i) in vitro functional validation of SLC6A14 in PC cell lines; (ii) targeted metabolomic profiling and coculture assays to validate the proposed metabolic–immune crosstalk; (iii) in vivo evaluation of genistein in PC xenograft models; and (iv) prospective clinical cohort validation of the five‐gene RF diagnostic model. Author Contributions Xing Liu and Jianghong Ou conceived and designed the study. Xing Liu drafted the manuscript. Yibin Li and Fangfang Qin performed parts of experiments and provided technical assistance. Xing Liu and Jianghong Ou supervised the project and revised the manuscript. All authors read and approved the final manuscript. Funding This work has been financed by the National Natural Science Foundation of China (82304796) and the HSTD Natural Science Foundation of Hunan Provincial (2026JJ82019). Ethics Statement The authors have nothing to report. Consent The authors have nothing to report. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Conditions of qPCR and primer sequences. Data S2: Biological basis of the red module. Data S3: OPLS‐DA and VIP analysis of metabolites in pancreatic cancer and control groups. (A) OPLS‐DA model diagnostics including the score plot, permutation test (pR2Y=0.001, pQ2=0.001), and observation diagnostics. The model validation parameters were R2X=0.306, R2Y=0.965, and Q2Y=0.954, indicating high predictive reliability (RMSEE = 0.086). (B) Variable importance in projection (VIP) plot derived from the OPLS‐DA model. Data S4: Graphs presenting the results of single‐cell sequencing data analysis. (A) Quality control chart for single‐cell sequencing data. (B) Graph showing the correlation between sequencing depth and the number of genes (left graph) (0.87 indicates a strong correlation), and the graph on the right shows the correlation between sequencing depth and the number of mitochondrial genes (right graph) (0.02 indicates almost no correlation). (C) Volcano plot of differentially expressed genes in single‐cell sequencing analysis. (D) Scatter plot of principal component analysis in single‐cell sequencing analysis. (E) Correlation diagrams of the 4 different principal components in single‐cell analysis with genes. (F) Heatmap of the 4 principal component analyses in single‐cell analysis. Data S5: Molecular docking results for all 19 candidate compounds targeting proteins. Data S6: Confusion matrix‐derived performance metrics of the random forest diagnostic model across training and validation cohorts. Acknowledgments We thank the GEO database for sharing the data. We thank the Home for Researchers editorial team (www.home‐for‐researchers.com) for language editing services. Data Availability Statement The original contributions presented in the study are included in the article/supporting information. Further inquiries can be directed to the corresponding authors. References Associated Data Supplementary Materials Table S1: Conditions of qPCR and primer sequences. Data S2: Biological basis of the red module. Data S3: OPLS‐DA and VIP analysis of metabolites in pancreatic cancer and control groups. (A) OPLS‐DA model diagnostics including the score plot, permutation test (pR2Y=0.001, pQ2=0.001), and observation diagnostics. The model validation parameters were R2X=0.306, R2Y=0.965, and Q2Y=0.954, indicating high predictive reliability (RMSEE = 0.086). (B) Variable importance in projection (VIP) plot derived from the OPLS‐DA model. Data S4: Graphs presenting the results of single‐cell sequencing data analysis. (A) Quality control chart for single‐cell sequencing data. (B) Graph showing the correlation between sequencing depth and the number of genes (left graph) (0.87 indicates a strong correlation), and the graph on the right shows the correlation between sequencing depth and the number of mitochondrial genes (right graph) (0.02 indicates almost no correlation). (C) Volcano plot of differentially expressed genes in single‐cell sequencing analysis. (D) Scatter plot of principal component analysis in single‐cell sequencing analysis. (E) Correlation diagrams of the 4 different principal components in single‐cell analysis with genes. (F) Heatmap of the 4 principal component analyses in single‐cell analysis. Data S5: Molecular docking results for all 19 candidate compounds targeting proteins. Data S6: Confusion matrix‐derived performance metrics of the random forest diagnostic model across training and validation cohorts. Data Availability Statement The original contributions presented in the study are included in the article/supporting information. Further inquiries can be directed to the corresponding authors.