A fatty acid metabolism–based deep learning model predicts biochemical recurrence and identifies NUDT19 as a candidate metabolic factor in prostate cancer Authors: Chen Wang, Li-Jing Zhu, Yan Zhang, Xiao-Fen Wu, Su-Qin Lei, Sheng-Ping Hu, Ming-Jie Sun, Qin-Zhou Yu, Ying Zhou, Jie Li Published: 2026 ABSTRACT PurposeTo develop a fatty acid metabolism–based deep learning model for predicting biochemical recurrence (BCR) in prostate cancer (PCa) and to identify recurrence-associated metabolic regulators.MethodsTranscriptomic data from TCGA and GEO GSE70769 were integrated to identify fatty acid metabolism–related genes and construct a deep learning model for BCR prediction. Tumor-infiltrating lymphocytes (TILs) were quantified from H&E-stained slides using a convolutional neural network–based approach. Single-cell RNA sequencing data were analyzed to identify candidate metabolic regulators enriched in malignant epithelial cells. Immunohistochemistry was performed to examine the protein expression patterns of NUDT19 and its expression associations with key fatty acid metabolism enzymes, including FASN, ACACA, and CPT1A. Functional roles were further evaluated using in vitro assays, xenograft models, and serum metabolomics.ResultsThe model effectively stratified PCa patients into high- and low-risk groups with distinct BCR-free survival outcomes. The high-risk group showed increased TIL infiltration, suggesting a more immune-infiltrated or inflammatory tumor microenvironment. Integrated single-cell and bulk transcriptomic analyses identified NUDT19 as a candidate metabolic regulator predominantly expressed in malignant epithelial cells and positively correlated with FASN, ACACA, and CPT1A. NUDT19 knockdown suppressed PCa cell proliferation, migration, and invasion, induced apoptosis, and inhibited tumor growth in vivo. Serum metabolomics further revealed that differential metabolites were enriched in fatty acid metabolism–related pathways.ConclusionsThis study establishes a fatty acid metabolism–based deep learning model for BCR prediction and identifies NUDT19 as a candidate metabolic regulator associated with lipid metabolic remodeling and PCa progression. These findings suggest a metabolically active, inflammation-associated recurrence subtype and support further investigation of NUDT19 as a potential therapeutic target in PCa.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12957-026-04446-9. Introduction Prostate cancer (PCa) is one of the most common malignancies in men and remains a leading cause of cancer-related mortality worldwide. Although localized PCa generally has a favorable prognosis, advanced, metastatic, and treatment-resistant disease continues to pose major clinical challenges. Current diagnostic and risk assessment strategies mainly rely on prostate-specific antigen (PSA), digital rectal examination, prostate biopsy, multiparametric magnetic resonance imaging, and clinicopathological parameters. However, these conventional indicators incompletely capture the molecular heterogeneity of PCa and have limited accuracy in predicting biochemical recurrence (BCR) and disease progression [1, 2]. Therefore, identifying robust molecular features associated with recurrence and malignant progression is essential for improving risk stratification and personalized management. PCa progression is driven by complex interactions among genetic alterations, androgen receptor (AR) signaling reprogramming, lineage plasticity, tumor microenvironment remodeling, and metabolic adaptation. Alterations in FOXA1, PTEN, TP53, and dysregulated AR signaling are key molecular events in disease progression. Although androgen deprivation therapy (ADT) and AR pathway inhibitors are standard treatments for advanced PCa, most patients eventually develop castration-resistant prostate cancer (CRPC), which is associated with reduced AR dependence, non-coding RNA dysregulation, and cell-cycle activation [3, 4]. For example, circPPFIA2 promotes PCa progression and enzalutamide resistance through the miR-646/miR-1200–ETS1 axis [4], whereas CDC20, DTL, and RRM2, as RB1/E2F1 axis–related factors, have been proposed as potential therapeutic targets in advanced PCa [5]. Under sustained therapeutic pressure, some tumors further undergo lineage reprogramming into AR-low or AR-negative neuroendocrine prostate cancer (NEPC), characterized by low PSA expression, high aggressiveness, and profound therapeutic resistance. MUC1/MUC1-C promotes stemness, epithelial–mesenchymal transition, and neuroendocrine phenotypes under AR-suppressed conditions [6], whereas MYCN contributes to endoplasmic reticulum stress, calcium dysregulation, and metabolic adaptation in NEPC [7]. Among these biological processes, metabolic reprogramming—particularly dysregulated lipid and fatty acid metabolism—has emerged as a critical hallmark of PCa progression, recurrence, and therapeutic resistance. SREBF1-mediated lipid and cholesterol remodeling enhances ferroptosis resistance and is associated with recurrence risk [3], while SLC25A10 promotes PCa cell proliferation, migration, and invasion by suppressing ferroptosis through the p62/KEAP1/Nrf2 axis and is linked to poor prognosis [2]. In parallel, the tumor microenvironment further shapes PCa metabolism and invasiveness. Cancer-associated fibroblasts facilitate tumor progression through extracellular matrix remodeling, inflammatory regulation, immune suppression, and metabolic support [8], and elevated CD146/MCAM in exosomes derived from aggressive PCa cells suggests potential value for disease monitoring and liquid biopsy [9]. In addition, SOX2 and RAB26 are implicated in tumor stemness, epithelial–mesenchymal transition, lineage plasticity, and treatment resistance [1, 10]. Nevertheless, current BCR prediction models are still largely based on clinicopathological variables or single-omics data and may fail to capture the metabolic heterogeneity and cellular-state diversity of PCa. In this study, we integrated transcriptomic data from TCGA-PRAD and the GEO dataset GSE70769 to develop a fatty acid metabolism–based deep learning model for BCR prediction in PCa. By combining bulk transcriptomics, single-cell RNA sequencing, immunohistochemistry, functional assays, xenograft models, and serum metabolomics, we identified NUDT19 as a candidate metabolic regulator predominantly expressed in malignant epithelial cells and associated with key fatty acid metabolism enzymes, including FASN, ACACA, and CPT1A. Our findings establish an integrated framework linking fatty acid metabolic remodeling, immune-infiltrated tumor microenvironment, and BCR risk, providing new insights into recurrence prediction and metabolism-targeted intervention for high-risk PCa. Materials and methods Data acquisition We downloaded mRNA sequencing expression data for 497 prostate cancer tissue samples and 51 normal prostate tissue samples from the TCGA database, along with corresponding clinical information, including patient age, TNM stage, Gleason score, BCR status, and BCR-free survival time. In addition, gene expression data and clinical information for the GSE70769 cohort, comprising 92 patients, were obtained from the GEO database. Three fatty acid metabolism–related gene sets—“Reactome Fatty Acid Metabolism,” “KEGG Fatty Acid Metabolism,” and “Hallmark Fatty Acid Metabolism”—were downloaded from the Molecular Signatures Database (MSigDB; http://www.gsea-msigdb.org/). We also searched GeneCards (https://www.genecards.org/) using the keyword “fatty acid metabolism” and retrieved genes with a relevance score > 50. After removing duplicates, a total of 331 fatty acid metabolism–related genes were collected for subsequent analyses. Immunohistochemical (IHC) specimens Postoperative paraffin-embedded tissue sections used for immunohistochemical staining were obtained from four patients with prostate cancer, including paired prostate cancer tissues and corresponding adjacent non-cancerous tissues from each patient. All postoperative pathological diagnoses were confirmed as prostate cancer. Written informed consent was obtained from all patients and/or their legal guardians. This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Lishui Central Hospital (Approval No. LSCHEC-2024-348). Differential gene identification and functional enrichment analysis Using R software, fragments per kilobase of transcript per million mapped reads (FPKM) values from the TCGA-PRAD cohort were converted to transcripts per million (TPM) values and subsequently log2-transformed for downstream analyses. For the GSE70769 dataset, raw microarray data were processed using the affy package, including background correction, normalization, and log2 transformation. Differential expression analysis of the 331 fatty acid metabolism–related genes in the TCGA-PRAD cohort was performed using the limma package, with the screening criteria set as |log2 fold change (FC)| > 0.585 and false discovery rate (FDR) < 0.05. Differentially expressed genes (DEGs) were identified by comparing 497 prostate cancer samples with 51 normal prostate tissue samples. The results were visualized using a volcano plot, and functional enrichment analysis was conducted using the Metascape online platform (https://metascape.org/). Development and validation of the deep learning–based prediction model First, univariate Cox proportional hazards regression analysis was performed in the TCGA-PRAD cohort to identify fatty acid metabolism–related genes significantly associated with biochemical recurrence–free survival (BCR-free survival). Genes with Cox regression P values < 0.05 were considered prognostically relevant and selected as candidate genes, including both protective and risk-associated genes. Before model construction, all gene expression data were log₂-transformed and z-score normalized to ensure comparability across samples and improve model convergence. Based on the identified prognostic genes, a deep learning–based neural network model was developed and implemented using the PyTorch framework [11]. The model stratified patients into high- and low-risk groups according to the predicted recurrence risk scores. To further evaluate survival differences, the survminer and survival R packages were used to stratify patients into high- and low-expression groups based on the expression levels of the model genes. Differences in BCR-free survival between groups were assessed, and Kaplan–Meier survival curves were generated. Furthermore, to enhance model interpretability, the SHapley Additive exPlanations (SHAP) framework was applied to quantify the contribution of each gene to model predictions. Visualization of SHAP values enabled the identification of key genes with the greatest influence on recurrence prediction, thereby providing biologically meaningful and interpretable insights at both the cohort and individual patient levels. Univariate and multivariate Cox proportional hazards regression analyses were performed to identify and validate independent prognostic risk factors. The distribution of patients in the high- and low-risk groups was further evaluated using three-dimensional principal component analysis (3D PCA). Histopathological image analysis and quantification of tumor-infiltrating lymphocytes Based on H&E-stained prostate cancer histopathological images obtained from the TCGA database, a dual convolutional neural network (CNN) framework [12] was applied to quantitatively assess tumor-infiltrating lymphocytes (TILs) while excluding necrotic regions. The lymphocyte detection network consisted of an 18-layer semi-supervised CNN initialized using a convolutional autoencoder (CAE). Image patches of 100 × 100 pixels were classified as lymphocyte-infiltrated or non-infiltrated using a sigmoid activation function and log-likelihood loss. During training, batch sizes were set to 32 for the CAE and 100 for the CNN, with learning rates of 0.03 and 0.001, respectively. The momentum parameters were set to 0.9 and 0.985, respectively, and the model was trained for 64 epochs. To improve model generalizability, data augmentation strategies, including random cropping, color perturbation, rotation, and mirroring, were applied. All networks were implemented using the Theano framework and trained on an NVIDIA Tesla K40 GPU. In addition, a necrosis segmentation network, Necrosis CNN, based on the DeconvNet architecture was used for pixel-wise segmentation of 500 × 500 μm² tissue regions, enabling accurate identification and exclusion of necrotic areas. By integrating the outputs of both CNNs, the spatial density and organizational patterns of TILs across prostate cancer tissues were systematically characterized. A spatial TIL infiltration score was subsequently calculated to quantify immune heterogeneity and compare immune infiltration patterns between the high- and low-risk groups. All CNN-based analyses were independently repeated three times under identical conditions to ensure robustness of the results. Single-cell, clinicopathological, and immunohistochemical analyses Using the SCAR online single-cell transcriptomic database, the SCAR_Atlas_0826 dataset was selected to identify metabolic regulatory factors enriched in malignant prostate cancer epithelial cells. The TIMER2 platform was then used to analyze the correlations between target genes and tumor purity, as well as key fatty acid metabolism–related enzymes. To assess the clinical relevance of NUDT19, its expression was compared across clinicopathological subgroups stratified by Gleason score and pathological T stage. Patients were grouped by Gleason score (≤ 7 vs. >7) and pathological T stage (T2, T3, and T4), and intergroup differences in NUDT19 expression were evaluated. Gene set enrichment analysis (GSEA) was performed to identify pathways enriched in the high-NUDT19-expression phenotype. In addition, immunohistochemistry (IHC) was conducted using tumor tissues and paired adjacent normal tissues from four prostate cancer patients to examine the protein expression patterns of NUDT19 and its expression associations with key fatty acid metabolism enzymes, including FASN, ACACA, and CPT1A. Cell lines and culture conditions The prostate cancer cell lines LNCaP, PC-3, DU145, and 22Rv1, as well as the normal prostate epithelial cell line RWPE-1, were obtained from ATCC (Shanghai, China). LNCaP, PC-3, DU145, and 22Rv1 cells were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin–streptomycin. RWPE-1 cells were maintained in Keratinocyte Serum-Free Medium (K-SFM) supplemented with epidermal growth factor (EGF) and bovine pituitary extract (BPE) according to the manufacturer’s instructions. All cells were cultured at 37 °C in a humidified atmosphere containing 5% CO₂. siRNA- and shRNA-mediated NUDT19 knockdown To silence NUDT19, small interfering RNAs (siRNAs) targeting NUDT19 were designed and synthesized by Hangzhou Molecular Diagnostics Biotechnology Co., Ltd. (Hangzhou, China). Transfection was performed using Lipofectamine 2000 (Thermo Fisher Scientific, USA; Cat. No. 11668030) according to the manufacturer’s instructions. Briefly, siRNA–Lipofectamine complexes were prepared in serum-free medium and added to prostate cancer cells, which were then incubated at 37 °C in a humidified atmosphere containing 5% CO₂. Cells were harvested 24 h post-transfection for downstream analyses, such as Western blotting, to evaluate the efficiency of transient NUDT19 knockdown. Based on the validated siRNA sequences, corresponding short hairpin RNA (shRNA) expression vectors targeting NUDT19 were subsequently constructed and cloned into lentiviral vectors to achieve stable gene silencing. Lentiviral particles were produced and used to infect prostate cancer cells, followed by puromycin selection to establish stable NUDT19 knockdown cell lines for subsequent in vivo experiments. Western blotting Total cell lysates were prepared using RIPA lysis buffer. Protein concentrations were determined using a BCA Protein Assay Kit (20201ES76, Yeasen Biotechnology, Shanghai, China). Equal amounts of protein were separated by 10% SDS-PAGE and transferred onto PVDF membranes. The membranes were blocked with 5% skim milk, incubated overnight with primary antibodies at 4 °C, washed with PBS, and then incubated with secondary antibodies at room temperature for 1 h. Protein bands were visualized using an enhanced chemiluminescence detection system. In vitro wound healing assay When cells reached 100% confluence, wounds were generated using a 200-µL pipette tip to create two parallel scratches. The cells were washed with PBS and cultured in serum-free medium for 48 h. Images were captured at 0 and 48 h using an inverted microscope. In vitro migration and invasion assays Migration and invasion assays were performed using 8-µm pore polycarbonate membrane chambers (BD Falcon, USA). For invasion assays, 2 × 10⁴ cells were seeded into the upper chambers coated with Matrigel, while the lower chambers were filled with 700 µL of medium containing 10% FBS. After 48 h, cells that had migrated or invaded through the membrane were stained with crystal violet and counted in three randomly selected fields under a microscope. Migration assays were performed using the same procedure without Matrigel coating. Cell viability assay Cell viability was evaluated using the Cell Counting Kit-8 (CCK-8) assay. Transfected cells were seeded into 96-well plates at a density of 5 × 10³ cells per well, with five replicate wells per group, and cultured at 37 °C. Cell viability was assessed by measuring absorbance at 450 nm at 24, 48, and 72 h after seeding. All experiments were independently performed in triplicate. Colony formation assay PC-3 and DU145 cells in the logarithmic growth phase were collected and seeded into 6-well plates at a density of 1,000 cells per well. Cells were cultured under standard conditions for 2 weeks to allow colony formation. Colonies were then fixed with 4% paraformaldehyde for 20 min and stained with 0.1% crystal violet for 20 min. Visible colonies were photographed and counted for quantitative analysis. Apoptosis analysis Cell apoptosis was analyzed by flow cytometry using Annexin V–FITC/PI staining. Briefly, 1–5 × 10⁵ cells were collected, washed twice with PBS, and resuspended in 100 µL of binding buffer. The cells were then incubated with 5 µL of Annexin V–FITC and 5 µL of propidium iodide (PI) solution for 10–15 min at room temperature in the dark. Apoptotic cells were subsequently detected by flow cytometry, and the data were analyzed using FlowJo software. Animal models All animal experiments were approved by the Animal Ethics Committee of the Hangzhou Institute of Medicine, Chinese Academy of Sciences (Approval No. AP2025‑04‑0504). Male nude mice were obtained from the Zhejiang Experimental Animal Center. For subcutaneous xenograft establishment, 3.5 × 10⁶ PC-3 prostate cancer cells were suspended in 100 µL of PBS containing 50% Matrigel and injected into the right axillary region of each mouse. Animals were randomly assigned to two groups (n = 8 per group): NUDT19 knockdown (shNUDT19) or negative control (shCtrl). Tumor volume and body weight were measured every other day. Mice were euthanized on day 42 post-inoculation or earlier if humane endpoints were reached, including excessive tumor burden or signs of distress. Excised tumors were photographed, weighed, and processed for subsequent analyses. Serum metabolomics analysis This study was approved by the Ethics Committee of Lishui Central Hospital, The Fifth Affiliated Hospital of Wenzhou Medical University (Approval No. LSCHEC-2024-348), and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants before enrollment. Based on predefined inclusion and exclusion criteria, a total of 90 participants were included, comprising 30 patients with prostate cancer (PCa) and 60 non-PCa controls. Clinical, imaging, and pathological data were collected for all participants, including age, PSA-related parameters [total PSA (tPSA) and free-to-total PSA ratio (f/tPSA)], prostate volume (PV), PSA density (PSAD), and multimodal imaging parameters, including PI-RADS score, elastography, and contrast-enhanced ultrasound findings. Peripheral venous blood samples were collected from all participants after overnight fasting using serum collection tubes without anticoagulant. Blood samples were allowed to clot at room temperature and then centrifuged to separate serum. The serum samples were aliquoted and stored at − 80 °C until analysis. Metabolites were extracted using a methanol-based protein precipitation method. The resulting supernatants were dried and subjected to ultra-high-performance liquid chromatography–tandem mass spectrometry (UHPLC–MS/MS) analysis in both positive and negative ion modes. Pooled quality control (QC) samples were prepared and analyzed to monitor system stability. Raw metabolomics data were processed through QC filtering, batch-effect correction, and total peak area normalization. Differential metabolites were identified using a two-sided Wilcoxon rank-sum test with Benjamini–Hochberg correction, with significance defined as FDR < 0.05 and |log₂FC| > 0.585. Fuzzy c-means clustering was subsequently performed using the Mfuzz package, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was conducted using the clusterProfiler package, with FDR < 0.05 considered statistically significant. Statistical analysis Data processing and statistical analyses were performed using R software, version 4.2.2. All experiments were performed in triplicate unless otherwise specified. Clinical correlation analyses were conducted using the Wilcoxon rank-sum test or the Kruskal–Wallis rank-sum test, as appropriate. Survival analysis was performed using the Kaplan–Meier method with the log-rank test. The predictive performance of the prognostic model was evaluated using receiver operating characteristic (ROC) curves. A two-sided P value < 0.05 was considered statistically significant. Results Differential expression analysis and prognostic model development Based on the TCGA-PRAD cohort, differential expression analysis was performed for 331 fatty acid metabolism–related genes using the criteria |log₂FC| > 0.585 and false discovery rate (FDR) < 0.05. A total of 497 prostate cancer samples and 51 normal tissue samples were compared, leading to the identification of 87 differentially expressed genes (DEGs), including 31 significantly upregulated and 56 significantly downregulated genes in prostate cancer tissues (Fig. 1A). Functional and pathway enrichment analyses of these DEGs were subsequently performed using the Metascape online platform, revealing significant enrichment in fatty acid metabolism–related biological processes and signaling pathways (Fig. 1B). Fig. 1Identification of fatty acid metabolism–related genes and construction of a prognostic model. A Volcano plot of differentially expressed fatty acid metabolism–related genes between prostate cancer and normal tissues in the TCGA-PRAD cohort. B Enrichment network showing functional and pathway enrichment of these genes, emphasizing fatty acid metabolism–related processes. C Overview of the PyTorch-based deep learning prognostic model Identification of fatty acid metabolism–related genes and construction of a prognostic model. A Volcano plot of differentially expressed fatty acid metabolism–related genes between prostate cancer and normal tissues in the TCGA-PRAD cohort. B Enrichment network showing functional and pathway enrichment of these genes, emphasizing fatty acid metabolism–related processes. C Overview of the PyTorch-based deep learning prognostic model Based on the differentially expressed fatty acid metabolism–related genes identified in the TCGA-PRAD cohort, univariate Cox regression analysis was further performed, identifying 11 prognostic genes significantly associated with BCR-free survival in prostate cancer (P < 0.05) (Supplementary Table 1). Using these prognostically relevant genes, a deep learning–based neural network model was constructed with PyTorch (Fig. 1C). The model demonstrated favorable predictive performance in the TCGA-PRAD cohort and was further validated in the independent GSE70769 cohort, in which patients in the high-risk group exhibited significantly shorter BCR-free survival than those in the low-risk group (P < 0.05). In addition, time-dependent receiver operating characteristic (ROC) curve analysis showed that the model achieved stable and reliable predictive performance at 1-, 3-, and 5-year time points, with area under the curve (AUC) values exceeding 0.7 in both the TCGA-PRAD and GSE70769 cohorts (Figs. 2A, B). To enhance model interpretability, SHapley Additive exPlanations (SHAP) analysis was performed and identified NUDT19, ADSL, and SDHA as the most influential genes contributing to the model predictions (Fig. 2C). Fig. 2Risk stratification in the TCGA-PRAD and GSE70769 cohorts and gene contributions to model prediction. A, B BCR-free survival curves and time-dependent ROC curves evaluating model performance in both cohorts. C SHAP analysis showing the relative importance of individual genes in predicting biochemical recurrence Risk stratification in the TCGA-PRAD and GSE70769 cohorts and gene contributions to model prediction. A, B BCR-free survival curves and time-dependent ROC curves evaluating model performance in both cohorts. C SHAP analysis showing the relative importance of individual genes in predicting biochemical recurrence Cox regression analysis and CNN-based TILs assessment Univariate and multivariate Cox regression analyses were performed in the TCGA-PRAD cohort. Univariate analysis showed that Gleason score, T stage, and risk score were significantly associated with BCR-free survival. Further multivariate analysis confirmed that T stage and risk score remained independent prognostic predictors of prostate cancer outcomes (Figs. 3A, B). In addition, three-dimensional principal component analysis (3D PCA) showed a clear separation between the high- and low-risk groups in the TCGA-PRAD cohort (Fig. 3C). Fig. 3Characteristic analysis of high- and low-risk groups in the TCGA-PRAD cohort. A Univariate Cox regression analysis in the TCGA-PRAD cohort. B Multivariate Cox regression analysis in the TCGA-PRAD cohort. C 3D principal component analysis plot of high- and low-risk groups in the TCGA-PRAD cohort. D, E H&E staining combined with CNN-based analysis demonstrated pronounced differences in the spatial distribution patterns of tumor-infiltrating lymphocytes (TILs) between high- and low-risk prostate cancer groups Characteristic analysis of high- and low-risk groups in the TCGA-PRAD cohort. A Univariate Cox regression analysis in the TCGA-PRAD cohort. B Multivariate Cox regression analysis in the TCGA-PRAD cohort. C 3D principal component analysis plot of high- and low-risk groups in the TCGA-PRAD cohort. D, E H&E staining combined with CNN-based analysis demonstrated pronounced differences in the spatial distribution patterns of tumor-infiltrating lymphocytes (TILs) between high- and low-risk prostate cancer groups In-depth analysis of hematoxylin and eosin (H&E)-stained prostate cancer tissue sections from the TCGA database revealed marked differences between the high- and low-risk groups in the spatial distribution, infiltration density, and structural organization of tumor-infiltrating lymphocytes (TILs). Convolutional neural network (CNN)-based TIL mapping demonstrated that the high-risk group had significantly higher TIL infiltration scores within tumor regions than the low-risk group, which showed relatively lower levels of TIL infiltration (P < 0.05) (Figs. 3D, E). NUDT19 is associated with lipid metabolism and disease progression in prostate cancer Single-cell transcriptomic analysis of the SCAR_Atlas_0826 prostate cancer dataset identified eight annotated cell types. NUDT19, the top contributor to the biochemical recurrence prediction model based on SHAP feature importance, was predominantly expressed in malignant epithelial tumor cells (Fig. 4A). Fig. 4Multi-omics analyses reveal lipid metabolism–associated features of NUDT19 in prostate cancer. A Single-cell transcriptomic analysis of the SCAR_Atlas_0826 dataset showing that NUDT19 is predominantly expressed in malignant epithelial cells. B Partial correlation analysis based on TCGA prostate cancer data, adjusted for tumor purity, showing positive correlations between NUDT19 and fatty acid metabolism–related enzymes (FASN, ACACA, and CPT1A). C Gene set enrichment analysis (GSEA) demonstrating enrichment of lipid metabolism–related pathways associated with high NUDT19 expression Multi-omics analyses reveal lipid metabolism–associated features of NUDT19 in prostate cancer. A Single-cell transcriptomic analysis of the SCAR_Atlas_0826 dataset showing that NUDT19 is predominantly expressed in malignant epithelial cells. B Partial correlation analysis based on TCGA prostate cancer data, adjusted for tumor purity, showing positive correlations between NUDT19 and fatty acid metabolism–related enzymes (FASN, ACACA, and CPT1A). C Gene set enrichment analysis (GSEA) demonstrating enrichment of lipid metabolism–related pathways associated with high NUDT19 expression We next evaluated the clinical relevance of NUDT19. As shown in Supplementary Fig. 1, NUDT19 expression was significantly higher in tumors with Gleason score > 7 than in those with Gleason score ≤ 7 (P = 0.048). NUDT19 expression also increased with advancing pathological T stage. Compared with T2 tumors, T3 and T4 tumors showed significantly higher NUDT19 expression (T3 vs. T2, P = 0.012; T4 vs. T2, P = 0.031). Although NUDT19 expression was higher in T4 than in T3 tumors, the difference was not significant (P = 0.090). These results suggest that elevated NUDT19 expression is associated with more advanced clinicopathological features. Partial correlation analysis using TCGA prostate cancer transcriptomic data, adjusted for tumor purity, showed that NUDT19 expression was positively correlated with key fatty acid metabolism enzymes, including the lipogenic enzymes FASN and ACACA (ACC1), and CPT1A, the rate-limiting enzyme of fatty acid β-oxidation (Fig. 4B). These findings suggest a potential link between NUDT19 and lipid metabolic reprogramming. GSEA further showed that high NUDT19 expression was associated with enrichment of lipid metabolism–related pathways. KEGG Peroxisome and KEGG Sphingolipid Metabolism were positively enriched, whereas KEGG Arachidonic Acid Metabolism was negatively enriched (Fig. 4C). IHC analysis comparing mean optical density values between normal prostate and PCa tissues showed higher expression of NUDT19, FASN, ACACA (ACC1), and CPT1A in PCa tissues, with concordant upregulation patterns of NUDT19 and these fatty acid metabolism–related enzymes (Fig. 5A, B). Fig. 5Immunohistochemical validation of NUDT19 and fatty acid metabolism–related enzymes in prostate cancer tissues. A Representative immunohistochemical (IHC) staining images of NUDT19, FASN, ACACA (ACC1), and CPT1A in normal and prostate cancer (PCa) tissues. B Quantification of mean optical density (Mean OD) values showing concordant upregulation of NUDT19 and fatty acid metabolism–related enzymes in PCa tissues Immunohistochemical validation of NUDT19 and fatty acid metabolism–related enzymes in prostate cancer tissues. A Representative immunohistochemical (IHC) staining images of NUDT19, FASN, ACACA (ACC1), and CPT1A in normal and prostate cancer (PCa) tissues. B Quantification of mean optical density (Mean OD) values showing concordant upregulation of NUDT19 and fatty acid metabolism–related enzymes in PCa tissues NUDT19 promotes prostate cancer progression in vitro and in vivo We first confirmed that NUDT19 expression was higher in prostate cancer cell lines than in the normal prostate epithelial cell line RWPE-1 (Fig. 6A). DU145 and PC-3 cells were subsequently selected for functional assays, and efficient NUDT19 knockdown was achieved in both cell lines through siRNA-mediated transient transfection (Fig. 6B). Fig. 6Expression of NUDT19 in prostate cancer cells and its role in regulating cell migration and invasion. A Western blot analysis of NUDT19 expression in prostate cancer cell lines LNCaP, PC-3, DU145, and 22Rv1 and the normal prostate epithelial cell line RWPE-1. B Western blot confirmed NUDT19 knockdown in DU145 and PC-3 cells. C Wound-healing assays showed a significant reduction in the migration rate in the knockdown group. D Transwell migration and invasion assays demonstrated significant inhibition of these processes in the knockdown group. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 Expression of NUDT19 in prostate cancer cells and its role in regulating cell migration and invasion. A Western blot analysis of NUDT19 expression in prostate cancer cell lines LNCaP, PC-3, DU145, and 22Rv1 and the normal prostate epithelial cell line RWPE-1. B Western blot confirmed NUDT19 knockdown in DU145 and PC-3 cells. C Wound-healing assays showed a significant reduction in the migration rate in the knockdown group. D Transwell migration and invasion assays demonstrated significant inhibition of these processes in the knockdown group. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 Functional assays demonstrated that NUDT19 silencing significantly suppressed the migratory and invasive capacities of prostate cancer cells. In the wound healing assay, NUDT19 knockdown markedly reduced the wound closure rate. Consistently, Transwell assays showed that the numbers of migrated and invaded cells were significantly decreased compared with the control group (Figs. 6C, D, P < 0.05). To further evaluate the role of NUDT19 in cell proliferation, CCK-8 assays were performed. NUDT19 knockdown significantly reduced cell viability at 24, 48, and 72 h compared with the control group (Fig. 7A, P < 0.05). Consistently, colony formation assays demonstrated that NUDT19 silencing significantly decreased both the number and size of colonies formed by DU145 and PC-3 cells (Fig. 7B, P < 0.05), indicating impaired proliferative capacity. Fig. 7NUDT19 knockdown suppresses proliferation, colony formation, and tumor growth while promoting apoptosis in prostate cancer cells. A CCK-8 assays showed that NUDT19 knockdown significantly reduced cell proliferation. B Colony formation assays demonstrated markedly fewer colonies following NUDT19 knockdown. C Flow cytometry analysis revealed a significantly higher proportion of apoptotic cells in NUDT19 knockdown cells. D Representative images of subcutaneous xenograft tumors showing decreased tumor volume and weight after NUDT19 knockdown. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 NUDT19 knockdown suppresses proliferation, colony formation, and tumor growth while promoting apoptosis in prostate cancer cells. A CCK-8 assays showed that NUDT19 knockdown significantly reduced cell proliferation. B Colony formation assays demonstrated markedly fewer colonies following NUDT19 knockdown. C Flow cytometry analysis revealed a significantly higher proportion of apoptotic cells in NUDT19 knockdown cells. D Representative images of subcutaneous xenograft tumors showing decreased tumor volume and weight after NUDT19 knockdown. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 Moreover, flow cytometric analysis using Annexin V–FITC/PI staining revealed that NUDT19 knockdown significantly increased the proportion of apoptotic cells compared with the control group (Fig. 7C, P < 0.05), suggesting that NUDT19 may play an important role in maintaining prostate cancer cell survival. To further validate these findings in vivo, a subcutaneous xenograft model was established using PC-3 cells. Compared with the control group (shCtrl), tumors derived from the NUDT19 knockdown group (shNUDT19) exhibited significantly reduced tumor volume and weight (Fig. 7D, all P < 0.05). These results indicate that NUDT19 promotes prostate cancer growth in vivo. Serum metabolomic profiling validates fatty acid metabolic reprogramming in prostate cancer To further validate the role of fatty acid metabolic reprogramming in PCa, we systematically analyzed serum metabolomic profiles from PCa patients and non-PCa individuals. The baseline clinical characteristics of all enrolled participants are summarized in Supplementary Table 2. Principal component analysis (PCA) revealed distinct serum metabolomic profiles between PCa samples and non-PCa controls, indicating substantial remodeling of the serum metabolome in PCa patients (Fig. 8A). Fig. 8Serum metabolomics reveals fatty acid metabolic reprogramming in PCa. A Principal component analysis (PCA) showed separation between PCa and non-PCa samples, suggesting distinct serum metabolomic profiles between the two groups. B Volcano plot of differential metabolites identified using the Wilcoxon rank-sum test with Benjamini–Hochberg correction, with thresholds of FDR < 0.05 and |log₂FC| > 0.585. C Fuzzy c-means clustering identified two major metabolite expression patterns, designated Clusters 1 and 2; lipid-related metabolites, such as LysoPC(20:5/0:0), were consistently elevated in the PCa group. D KEGG pathway enrichment analysis highlighted significant dysregulation of lipid metabolism–related pathways, particularly α-linolenic acid and linoleic acid metabolism Serum metabolomics reveals fatty acid metabolic reprogramming in PCa. A Principal component analysis (PCA) showed separation between PCa and non-PCa samples, suggesting distinct serum metabolomic profiles between the two groups. B Volcano plot of differential metabolites identified using the Wilcoxon rank-sum test with Benjamini–Hochberg correction, with thresholds of FDR < 0.05 and |log₂FC| > 0.585. C Fuzzy c-means clustering identified two major metabolite expression patterns, designated Clusters 1 and 2; lipid-related metabolites, such as LysoPC(20:5/0:0), were consistently elevated in the PCa group. D KEGG pathway enrichment analysis highlighted significant dysregulation of lipid metabolism–related pathways, particularly α-linolenic acid and linoleic acid metabolism Using the Wilcoxon rank-sum test with Benjamini–Hochberg correction and thresholds of FDR < 0.05 and |log₂FC| > 0.585, we identified 43 significantly altered metabolites between the two groups (Fig. 8B). These differential metabolites exhibited two major expression patterns, designated Clusters 1 and 2 (Fig. 8C). Notably, lipid-related metabolites in Cluster 1, such as LysoPC(20:5/0:0), were consistently increased in the PCa group, suggesting enhanced lipid metabolic activity. Further KEGG pathway enrichment analysis revealed significant perturbations in multiple lipid metabolism–associated pathways (Fig. 8D). In particular, alpha-linolenic acid and linoleic acid metabolism pathways were prominently enriched, consistent with the fatty acid metabolism–related gene model described above and further supporting a critical role for fatty acid metabolic reprogramming in PCa progression. Discussion In this study, we developed and validated a fatty acid metabolism–based deep learning model for predicting biochemical recurrence (BCR) in prostate cancer (PCa) and explored its biological implications through multidimensional analyses. By integrating bulk transcriptomics, interpretable deep learning, single-cell RNA sequencing, digital histopathology, functional experiments, xenograft validation, and serum metabolomics, we established a framework linking lipid metabolic dysregulation to recurrence-associated tumor behavior. Our findings suggest that fatty acid metabolism–related transcriptional programs provide prognostically informative signals beyond conventional clinicopathological variables and highlight NUDT19 as a candidate metabolic factor associated with aggressive PCa phenotypes. Metabolic reprogramming is a hallmark of cancer progression, and lipid metabolism is particularly relevant in PCa [13–15]. Unlike many malignancies that predominantly rely on glycolytic reprogramming [16, 17], PCa frequently exhibits enhanced lipid synthesis, fatty acid uptake, and fatty acid oxidation to support membrane biosynthesis, energy production, redox balance, and oncogenic signaling [18]. Consistent with this biological context, our model showed favorable predictive performance across independent cohorts, and the model-derived risk score remained independently associated with BCR after adjustment for established clinical parameters. These findings indicate that lipid metabolism–related molecular features may complement conventional clinicopathological factors, such as Gleason score and pathological stage, in postoperative recurrence risk stratification [19]. A key strength of this study is the use of SHAP analysis to improve the interpretability of the deep learning model. Although deep learning can capture complex nonlinear patterns in high-dimensional transcriptomic data, its biological interpretation remains challenging [20]. SHAP-based feature attribution identified NUDT19 as the strongest contributor to BCR prediction among the fatty acid metabolism–related genes. Higher NUDT19 expression was generally associated with positive SHAP values, suggesting that its upregulation contributed to an increased predicted recurrence risk. This finding provided a rationale for further experimental investigation of NUDT19 and helped link the predictive model to a biologically interpretable candidate factor. NUDT19 belongs to the Nudix hydrolase family, but its role in PCa remains poorly understood [21]. In this study, NUDT19 was expressed at higher levels in androgen-independent PCa cell lines than in androgen-sensitive LNCaP cells, suggesting a potential association with androgen-independent progression. Single-cell transcriptomic analysis further indicated that NUDT19 was mainly expressed in malignant epithelial cells and was associated with peroxisomal and sphingolipid metabolic pathways. These findings imply that NUDT19 may reflect tumor cell–intrinsic lipid metabolic remodeling rather than changes in stromal or immune composition. Given the roles of peroxisomal lipid metabolism and sphingolipid signaling in oxidative stress regulation, apoptosis, membrane dynamics, and cell motility, NUDT19 may participate in metabolic adaptation during aggressive PCa progression [22, 23]. However, its precise enzymatic substrates and downstream metabolic consequences require further biochemical validation. Functional experiments supported the biological relevance of NUDT19 in PCa. Knockdown of NUDT19 in DU145 and PC-3 cells inhibited proliferation, migration, invasion, and colony formation while promoting apoptosis. In xenograft models, NUDT19 silencing also suppressed tumor growth. These results suggest that NUDT19 is not merely a computationally identified prognostic feature but may also contribute functionally to malignant phenotypes, particularly in androgen-independent PCa models. Nevertheless, further studies using additional cellular systems, patient-derived organoids, rescue experiments, and mechanistic assays are needed to confirm the specificity and generalizability of these effects. NUDT19 knockdown reduced the migratory and invasive capacities of PCa cells, suggesting that NUDT19 may influence cellular programs involved in tumor dissemination, potentially through lipid metabolism–associated effects on migratory machinery and microenvironmental adaptation. One possible explanation is that NUDT19-associated lipid metabolic remodeling may affect extracellular matrix (ECM) remodeling and invasive behavior. MMP-mediated ECM degradation, particularly involving MMP1 and MMP2, is a well-established mechanism supporting local invasion and metastatic dissemination in multiple cancers, including PCa [24–27]. NUDT19 may also be indirectly associated with EMT-like phenotypic changes and pro-migratory signaling programs. For example, activation of TGF-β signaling and EMT-related factors, including Snail, ZEB1/2, N-cadherin, and Vimentin, can enhance tumor cell motility, invasiveness, and survival [28, 29]. In parallel, lipid metabolic remodeling may provide structural and signaling support for filopodia, lamellipodia, and invadopodia formation, whereas β1-integrin-mediated ECM stiffness sensing may coordinate cytoskeletal reorganization, MMP localization at the invasive front, and directional migration [30, 31]. Thus, NUDT19 may contribute to PCa invasion by influencing lipid metabolism–associated changes in membrane dynamics, cytoskeletal organization, ECM interaction, and pro-migratory signaling. However, these mechanisms remain inferential and were not directly tested in the present study. Future experiments assessing EMT markers, MMP1/MMP2 expression and activity, invasive protrusion dynamics, actin organization, and TGF-β/β1-integrin downstream signaling after NUDT19 perturbation are required to clarify how NUDT19 regulates PCa invasion. Serum metabolomic analysis identified LysoPC (20:5/0:0) as elevated in PCa patients compared with non-cancer controls. LysoPC species are bioactive lipid metabolites involved in membrane remodeling [32], inflammatory signaling [33], oxidative stress [34], and cell migration [35]. Elevated LysoPC (20:5/0:0) may reflect increased phospholipid turnover and altered membrane lipid metabolism in PCa, consistent with the metabolic requirements of proliferating tumor cells. In addition, LysoPCs may influence immune recruitment and inflammatory responses [36, 37], suggesting a potential link between lipid metabolic changes and immune contexture. However, this study did not establish a direct regulatory relationship between LysoPC (20:5/0:0) and NUDT19; therefore, their association should be interpreted as part of a broader lipid metabolic dysregulation program rather than as evidence of a confirmed causal axis. The immune microenvironment analysis indicated that fatty acid metabolism–related risk stratification may be associated with distinct immune features. CNN-based histopathological analysis showed higher tumor-infiltrating lymphocyte (TIL) infiltration scores in the high-risk group, indicating a more immune-infiltrated or inflammatory tumor contexture. However, bulk RNA-seq analysis of PDCD1 and LAG3 did not reveal significant differences between the risk groups (Supplementary Fig. 2; both P > 0.05), suggesting that classical exhaustion markers measured at the bulk-tissue level may not fully capture local immune dysfunction. Therefore, the high-risk subgroup may represent an immune-infiltrated but not necessarily immune-effective phenotype. Further single-cell RNA sequencing, spatial transcriptomics, or multiplex immunohistochemistry will be required to define the functional states and spatial organization of immune cells in this metabolically high-risk subtype. Compared with previous studies focusing mainly on individual lipid metabolism genes or conventional prognostic signatures [18, 38], our study integrates deep learning, model interpretability, single-cell transcriptomics, digital pathology, functional validation, and serum metabolomics to provide a more comprehensive view of recurrence-associated metabolic remodeling in PCa. Clinically, the fatty acid metabolism–based model may complement existing postoperative risk stratification systems and help identify patients with unfavorable molecular features who may require closer surveillance or consideration of adjuvant treatment. NUDT19 may serve as a candidate prognostic biomarker and potential therapeutic vulnerability, although standardized detection methods, prospective validation, and assessment of its incremental value over established nomograms are required before clinical translation. Several limitations should be acknowledged. First, this study was primarily based on retrospective public datasets, which may introduce selection bias, batch effects, and platform-related heterogeneity. Although independent validation cohorts were included, larger multicenter prospective cohorts are needed to confirm the robustness and generalizability of the model. Second, SHAP analysis identifies important predictive features but does not prove causality. Although functional experiments support a tumor-promoting role of NUDT19, its molecular mechanisms remain incompletely defined. Third, immune microenvironment analysis was based mainly on digital pathology and bulk RNA-seq, limiting the resolution of immune cell heterogeneity, spatial localization, and functional states. Fourth, the serum metabolomics cohort was relatively small, and the prognostic value of LysoPC (20:5/0:0) requires further validation in larger longitudinal cohorts. Finally, the proposed connections among NUDT19, lipid metabolic remodeling, invasion, and immune regulation remain partly inferential and should be tested through targeted mechanistic and interventional studies, which will be essential for determining the translational relevance of this fatty acid metabolism–associated recurrence framework. Conclusion In summary, this study supports an important role for fatty acid metabolic reprogramming in biochemical recurrence and aggressive tumor behavior in prostate cancer. We developed a fatty acid metabolism–based deep learning model that showed favorable prognostic performance in both the TCGA-PRAD cohort and an independent validation cohort. Integrative analyses identified NUDT19 as a candidate metabolic factor associated with lipid metabolism–related enzymes, advanced clinicopathological features, and increased recurrence risk. Functional experiments further demonstrated that NUDT19 promotes prostate cancer cell proliferation, migration, invasion, colony formation, and tumor growth. In addition, serum metabolomic and digital pathology analyses suggested that fatty acid metabolic dysregulation may be linked to systemic lipid metabolic alterations and distinct immune infiltration patterns. These findings highlight NUDT19 as a potential prognostic biomarker and candidate therapeutic vulnerability, although further mechanistic studies and prospective clinical validation are required before clinical translation. Supplementary Information Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2. Supplementary Material 3. Supplementary Material 3. Supplementary Material 4. Supplementary Material 4. Supplementary Material 5. Supplementary Material 5. Supplementary Material 6. Supplementary Material 6.