Machine learning identifies disulfidptosis-related gene signature for pancreatic cancer prognosis and immune infiltration

Abstract

Background

Pancreatic cancer (PC) is a leading cause of cancer-related mortality due to late diagnosis and limited treatments. Disulfidptosis, a novel form of regulated cell death, is implicated in disease pathogenesis. This study explores disulfidptosis-related gene (DRG) signatures to identify prognostic biomarkers and immune infiltration patterns in PC using bioinformatics and single-cell analyses.

Methods

Differential analysis of PC gene expression identified overlapping DRGs. Hub genes were selected via machine learning, and a prognostic model was built, validated with GSE28735. Single-cell RNA sequencing (scRNA-seq) from GSE212966 examined DRG expression in the tumor microenvironment.

Results

Among 42 DRGs, 19 prognostic genes were first identified by univariate Cox analysis. Using the R package Mime integrating 10 machine learning algorithms, and further evaluated by StepCox[both] + RSF, 8 hub genes (NDUFA11, MYH14, TRIP6, ACTN1, ANP32E, PML, CD2AP, RPN1) were extracted. An optimized disulfidptosis-based risk score model was constructed via StepCox[both] combined with the random survival forest (RSF) algorithm, achieving a concordance index (C-index) of 0.9 in the training set though external validation yielded moderate performance (C-index = 0.6), indicating potential for refinement. RPN1 and PML were significantly associated with survival and enriched in cell cycle and immune regulation pathways. scRNA-seq revealed RPN1 upregulation in macrophages, suggesting an immunosuppressive role. Immune infiltration and drug sensitivity analyses highlighted distinct profiles between high- and low-risk groups.

Conclusion

This study establishes a disulfidptosis-based prognostic model for PC, identifying RPN1 and PML as key biomarkers. These findings provide novel insights into PC prognosis and immune dynamics, supporting the development of targeted diagnostic and therapeutic strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04463-w.

Introduction

Pancreatic cancer (PC) is one of the most lethal solid malignancies, with a global 5-year survival rate under 10% and increasing incidence due to late diagnosis, early metastasis, therapeutic resistance, and an immunosuppressive tumor microenvironment (TME) [1], with radiologic-pathologic correlations of pancreatic lesions aiding in the characterization of PDAC and other malignancies, though early detection remains challenging [2]. KRAS mutation is a hallmark of pancreatic cancer, driving tumor progression and representing a key therapeutic target [3]. Oncogenic KRAS specifically triggers metabolic reprogramming in PDAC, exacerbating energy demands and resistance [4]. Although significant advances have been made in surgical, chemotherapeutic, targeted, and immunotherapeutic approaches, durable improvements in clinical outcomes remain limited [5], with immunotherapy challenges including side effects like hypophysitis causing abdominal symptoms [6], novel agents like cycloruthenated ruthenium complexes showing promise in overcoming multidrug resistance in various cancers [7]. Consequently, the identification of novel prognostic biomarkers and actionable targets is urgently needed to improve patient survival [8].

Recently, disulfidptosis, a programmed cell death triggered by disulfide stress and cytoskeletal collapse, has emerged as a tumor vulnerability in preclinical models [9]. Its regulation of metabolic stress responses and immune modulation offers promising translational potential [10, 11]. Analogous cell death mechanisms, such as ferroptosis, have been leveraged for prognostic signatures involving mRNAs and lncRNAs in gastric cancer [12]. However, integrative analyses of disulfidptosis-related gene (DRG) signatures in PC, linking them with prognosis and TME infiltration, are still lacking [13].

Machine learning (ML) models built on bulk transcriptomics have achieved high prognostic accuracy (C-index > 0.8) in PC—as evidenced by a recent radiological-clinical random survival forest model achieving C-index of 0.828 in training and validation cohorts [14]. Similar radiomics-based ML approaches have been applied to predict survival in other gastrointestinal cancers, such as colorectal liver metastases treated with hepatic arterial infusion chemotherapy [15]. Single-cell RNA sequencing (scRNA-seq) has revealed cellular heterogeneity including macrophage polarization and T-cell exhaustion in the PC TME [16–18]. Pan-cancer analyses leveraging bioinformatics, such as those identifying RUNX transcription factors as carcinogenic biomarkers, further demonstrate the power of integrative omics for uncovering prognostic signatures across tumor types [19].

Recent studies have identified disulfidptosis-related gene (DRG) signatures in various cancers, including gastric and colorectal cancers [20–22], with miRNA dysregulation axes like miR-144-5p/RNF187 also contributing to progression in colorectal cancer [23]. However, pancreatic cancer-specific DRG analyses are limited, with prior work focusing on bulk transcriptomics without single-cell validation [11, 24, 25].

Herein, we aimed to develop a ML-based prognostic signature using DRGs in PC; elucidate its association with immune infiltration patterns, tumor mutational burden, and therapy sensitivity; and validate key genes at single-cell resolution using scRNA-seq. This comprehensive framework fills a critical gap and supports precision oncology efforts by linking disulfidptosis pathways to TME dynamics and clinical outcomes in pancreatic cancer.

Pancreatic ductal adenocarcinoma (PDAC) signatures focusing on driver mutations such as KRAS, CDKN2A, TP53, and SMAD4 [26] or tumor microenvironment (TME)-related gene panels [27] have advanced our understanding of PDAC’s molecular complexity. However, these signatures often rely on bulk RNA sequencing, lack single-cell resolution, or show inconsistent external validation, limiting their prognostic reliability [28]. Emerging research on novel cell death mechanisms, such as disulfidptosis, highlights their potential for identifying therapeutic targets, yet few signatures incorporate these pathways [29]. To address these gaps, our disulfidptosis-based prognostic model integrates bulk RNA-seq (TCGA-PAAD, GTEx, GSE28735) and single-cell RNA-seq (GSE212966), enhancing predictive accuracy and clinical utility through comprehensive TME and genomic analyses.

Materials and methods

Data acquisition and pre-processing

The transcriptomic data of the TCGA-PAAD cohort, comprising 178 pancreatic tumor tissues and 4 normal pancreatic tissues, along with single nucleotide polymorphism (SNP), copy number variation (CNV), and matched clinical information, were obtained from The Cancer Genome Atlas (TCGA). The expression data (TPM) of normal pancreatic tissues from the GTEx database were downloaded via the UCSC Xena platform (https://xena.ucsc.edu/) and normalized via log2(x + 0.001) transformation, which is consistent with the GTEx format. By integrating 182 TCGA samples and 167 normal pancreatic samples from GTEx, a final dataset comprising 178 tumor samples and 171 normal samples was constructed, covering the expression profiles of 36,085 genes for downstream analyses. Principal component analysis (PCA) was applied to evaluate potential batch effects between the TCGA and GTEx data. DEGs were identified via the Wilcoxon rank-sum test, with selection thresholds of FDR < 0.05 and |log2(fold change)| > 0.585. Visualization of DEGs was performed using the R package “ggplot2”. DRDEGs were defined by intersecting DEGs, mRNA-encoding DEGs, and DRGs (Table S1). The GEO dataset GSE28735 was retrieved using the keywords “Pancreatic adenocarcinoma”, species set to “Homo sapiens”, and sample type restricted to “tissue”.

Enrichment analysis

To investigate the functional roles and pathway associations of the hub genes, a significance threshold of p < 0.05 was applied. Gene symbol IDs of each DRGs were converted to Entrez Gene IDs using the R package “org.Hs.eg.db”.

Subsequently, Gene Ontology (GO) enrichment analysis [30] and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis [31] were performed on the DEGs using the “clusterProfiler” package [32]. GO analysis was conducted across three categories: molecular function (MF), cellular component (CC), and biological process (BP). The results were visualized using the R packages “ggplot2” and “ComplexHeatmap” [33].

Construction and verification of the prognostic model

Univariate Cox regression analysis was conducted on the expression profiles of hub genes to assess their prognostic relevance. Genes with a hazard ratio (HR) ≠ 1 and FDR < 0.05 were selected as prognosis-associated candidates for downstream analysis. Using the “Mime” [34] package, we constructed an optimal prognostic model and identified feature genes. To identify hub genes, 42 DRGs were filtered by univariate Cox regression, yielding 19 prognostic candidates. These were further analyzed using the R package Mime, which integrates 10 machine learning algorithms for feature selection. Based on consensus ranking and performance in StepCox[both] + RSF models, 8 robust hub genes were obtained. Mime, a machine learning framework, integrates ten classic algorithms: RSF, Enet, StepCox, CoxBoost, plsRcox, superpc, GBM, survivalsvm, Ridge and Lasso. A total of 117 combinations were applied with K-fold cross-validation for model training. The model’s performance, assessed using the C-index, demonstrated its capability to stratify patients into high- and low-risk survival groups. OS (Overall survival) analysis was performed on the training cohort based on the risk scores and grouping information to construct the prognostic model, which was further validated in the independent testing cohort.

Clinicopathologic correlation, independent prognostic analysis and nomogram model construction

Clinical information and risk scores were integrated, and patients were stratified according to clinicopathological characteristics. The Wilcoxon and Kruskal–Wallis rank-sum tests were used to evaluate differences in risk scores across subgroups. To assess whether the risk score functions as an independent prognostic indicator for pancreatic cancer, both univariate and multivariate Cox regression analyses were performed. ROC (Receiver Operating Characteristic) curves were generated using the R package “timeROC” to evaluate the predictive accuracy and independence of the risk model in comparison with clinical traits. A prognostic nomogram was developed using the R package “survival”, incorporating clinical parameters such as age, tumor stage, pathological grade, sex, and the risk score. The performance of the nomogram was assessed through calibration curves to evaluate its predictive consistency [35].

Immune analysis

To assess alterations in the tumor microenvironment (TME) between the low- and high-risk groups, immune cell composition was estimated using eight computational algorithms implemented via the R packages “immunedeconv” [36] and “IOBR” [37]. Single-sample gene set enrichment analysis (ssGSEA) was conducted using the “GSVA” and “GSEABase” R packages to calculate infiltration scores for 16 immune cell types and activity scores for 13 immune-related pathways across individual patients. TIDE analysis was employed to simulate immune evasion mechanisms and predict sensitivity to immune checkpoint blockade (ICB). TIDE scores—including T cell dysfunction, exclusion, and microsatellite instability (MSI)—were calculated using the “tidepy” package [38, 39] in a Jupyter Notebook environment. The Cancer Immunome Atlas (TCIA) (https://tcia.at/) was used to predict the responsiveness of each sample to CTLA-4 or PD-1 inhibition, and comparative analysis was conducted between the two risk subgroups. Additionally, the Estimation of Systemic Immune Response (EaSIeR) [40] framework was utilized to predict ICB responsiveness, which incorporates comprehensive TME characteristics. A lower TIDE score and a higher EaSIeR score indicate a more favorable predicted response to immunotherapy. Immune cell infiltration estimates for all TCGA cancer types were retrieved from the Tumor Immune Estimation Resource (TIMER2.0, http://timer.comp-genomics.org/) [41].

Mutation analysis

Somatic mutation analysis was performed using the R package “maftools” [42]. TMB [43] was defined to include somatic mutations, insertions, base substitutions, coding mutations, deletions and was calculated by dividing the total number of mutations by 30 megabases (Mb). The association between the risk score and TMB was assessed using Spearman correlation analysis. Mutation profiles of the high- and low-risk groups were visualized using the “oncoplot” function, highlighting the top 20 most frequently mutated genes in each group, providing insights into potential genomic drivers of prognosis and their integration into the risk stratification model. Fisher’s exact test was employed to identify genes with significantly different mutation frequencies between the two risk groups.

Gene expression and distribution exploration

The Human Protein Atlas (HPA) database (https://www.proteinatlas.org/) [44] was utilized to explore protein expression patterns in normal cells, normal tissues, and cancerous tissues. To validate gene expression at the protein level, immunohistochemistry (IHC) images of normal pancreatic and pancreatic cancer tissues were retrieved from the HPA platform. Additionally, associated datasets and images were collected to examine the spatial distribution of gene expression across various subcellular compartments and distinct cell types.

Drug sensitivity analysis

To investigate differences in drug sensitivity between the high- and low-risk groups, the R package “oncopredict” [45] was applied. Gene expression profiles and drug sensitivity data for FDA-approved compounds were obtained from the CellMiner database (https://discover.nci.nih.gov/cellminer/) to further explore associations between therapeutic response and genes included in the predictive model. Spearman correlation analysis was used to evaluate the relationships between gene expression levels and drug responsiveness.

Single-cell RNA sequencing analysis

The single-cell RNA sequencing matrix of the GSE212966 dataset was processed using the R package “Seurat” [46]. The expression patterns of RPN1 and PML were visualized via the functions’FeaturePlot’, ‘DotPlot’, and ‘VlnPlot’. Pseudotime trajectory analysis was performed using the R package slingshot.

Statistical analysis

All analyses and R package implementations were carried out using R software (version 4.4.1), Python (version 3.9.13), and the Anaconda environment. The Wilcoxon rank-sum test or Student’s t-test was used for comparisons between two groups, whereas the Kruskal–Wallis test was used for comparisons among more than two groups. Pearson correlation analysis was performed for parametric variables, and Spearman correlation analysis was used for nonparametric variables. All the statistical tests were two-tailed, and p < 0.05 was considered statistically significant.

Results

Detection and analysis of differentially-expressed DRDEGs

As shown in Fig. 1, we described the workflow in this study. Transcriptomic data from TCGA (178 tumor, 4 normal samples) and GTEx (167 normal samples) were integrated, with PCA showing clear separation between tumor and normal tissues (Fig. 2a). We identified 6,143 DEGs (5,696 upregulated, 447 downregulated) (Fig. 2b, c), with 42 DRDEGs overlapping with mRNAs and 108 DRGs (Fig. 2d). Chromosomal mapping revealed CNV alterations in all 42 genes, most frequently in ZHX2, ARNT, and ANP32E (Fig. 2e, g). All DRDEGs presented significant expression differences between tumor and normal tissues (Fig. 2f). Among 172 tumor samples, 14 (8.14%) showed mutations, mainly missense types, with MYH14, CHD4, TLN2 and MYH10 being the most affected (Fig. 2h). GO and KEGG analyses revealed enrichment in cytoskeleton organization, energy metabolism, and immune processes (Fig. 2i, j; Table S2, S3). These findings offer initial insights into pancreatic cancer mechanisms and support future research directions.

Fig. 1.

Workflow of our study

Fig. 2.

Identification and analysis of differentially expressed disulfidptosis-related genes (DRDEGs). a Principal Component Analysis (PCA) plot of PAAD samples. The green dots represent normal samples, and the red dots represent tumor samples. b Volcano plot of differentially expressed genes in PAAD tissue compared with normal tissue. The thresholds were |log2(fold change)| > 0.585 and FDR < 0.05. c Heatmap of differentially expressed genes. d Venn diagram showing the intersection of differentially expressed genes, mRNAs and dithiol death-related genes. e Chromosomal locations of CNVs. f Boxplots showing the expression of hub genes in pancreatic cancer samples from the TCGA merged with the GTEx database. g Frequencies of CNV gains and losses. h Gene mutations of the hub genes. i GO enrichment analysis. j KEGG enrichment analysis

Establishment and validation of the DRDEGs-related prognostic model

To identify prognosis-related genes, univariate Cox regression analysis of 42 hub genes revealed that 19 genes were significantly linked to survival (p < 0.05) (Fig. 3a). Spearman correlation revealed positive links between RPN1-ACTN4 and PML-TRIP6 (Fig. S1). A total of 117 machine learning models were tested (Fig. 3b); the StepCox[both] + RSF model had the highest average C-index (0.75) (Fig. 3c). In the training cohort, the AUCs (Area Under the Curve) for 0.5-, 1-, 2-, and 3-year OS were 0.93, 0.95, 0.95 and 0.95, respectively. In the validation cohort, the corresponding AUCs were 0.67, 0.63, 0.68 and 0.77 (Fig. 3d). Risk scores were calculated per patient, and 39.53 (median from the training cohort) was used to define high- and low-risk groups. Most recurrences were observed in the high-risk group, which also presented distinct gene expression profiles (Fig. 3e). This trend has been validated externally. K–M analysis revealed that high-risk patients had significantly shorter OS (training: p < 0.0001; validation: p = 0.016) (Fig. 3f). RMST analysis confirmed a poorer prognosis: the 3-year OS in the training cohort was 359.22 vs. 999.06 days; in the validation cohort, it was 502.65 vs. 706.68 days (Fig. 3g).

Fig. 3.

Development of the prognostic model and survival analysis. a Forest plot of the univariate Cox regression analysis results showing hazard ratios for the19 selected genes. b C-index calculations for machine learning prediction model combinations in both training and validation sets. c Regression coefficients of 8 genes obtained from StepCox regression (left); changes in error rate of random survival forest model with increasing number of decision trees; importance scores of genes in the random survival forest model (right). d AUC values at 0.5-year,1-year, 2-year and 3-year for the optimal model StepCox[both] + RSF. e Rankdot plot and scatter plot showing risk score distribution and patient survival status; heatmap displaying expression levels of 19 selected prognostic genes in high-risk and low-risk populations. f Kaplan-Meier OS survival curves of the prognosticmodel in the training set (left) and validation set (right). g RMST of 3-year OS for both risk groups. The left panel shows the high-risk group and the right panel shows the low-risk group

Clinicopathologic correlation, independent prognostic analysis and nomogram model construction

Tumor staging is essential for guiding treatment and predicting outcomes. To assess how risk scores are relate to clinical characteristics, subgroup analyses were performed across age, sex, tumor grade, clinical stage, T stage (tumor infiltration), M stage (distant metastasis), and N stage (lymph node involvement). The risk scores did not significantly differ among the subgroups in terms of age, sex, M stage, or N stage (Fig. 4a, b, f, g). However, significant differences were found among tumor grades, including G1 vs. G2, G1 vs. G3, G2 vs. G3, and G3 vs. G4 (p < 0.05) (Fig. 4c). Both univariate and multivariate Cox regression identified the risk score and age as independent prognostic factors (Fig. 4h, i). Time-dependent ROC analysis revealed that the risk score had greater predictive accuracy for 0.5-, 1-, 2-, and 3-year OS than traditional clinical features such as stage and grade did (Fig. 4j–m). A prognostic nomogram incorporating clinical variables and the risk score was constructed (Fig. 4n), achieving a high C-index of 0.9166. The calibration plots demonstrated strong agreement between the predicted and actual survival rates (Fig. 4o). Finally, chi-square tests revealed no significant differences in clinicopathological features between the high- and low-risk groups, indicating that the model’s value is independent of traditional staging systems (Fig. S2).

Fig. 4.

Clinical-pathological correlations, independent prognostic analysis and nomogram model construction. a–g Boxplots showing risk scores (y-axis) against clinical characteristics (x-axis), including age, gender, tumor grade (Grade), clinical stage (Stage), tumor infiltration percentage (T), distant metastasis (M) and number of positive lymph nodes (N). h, i Univariate and multivariate Cox regression analyses of clinical indicators and risk scores with OS. j–m ROC curves for 0.5-, 1-, 2- and 3-year OS comparing risk scores with clinical characteristics (age, gender, Grade, Stage, T, M, N). n Nomogram integrating risk scores and key clinical characteristics. o Calibration plots of actual risk probabilities

Immune analysis

To investigate immune infiltration differences between the high- and low-risk groups, seven algorithms were used to assess the TME (Fig. 5a). The heatmap showed higher immune cell infiltration in high-risk patients, although ESTIMATE analysis revealed no significant differences in the immune score, stromal score, ESTIMATE score, or tumor purity (Fig. 5b). Single-sample GSEA (ssGSEA) of 29 immune-related gene sets revealed that the low-risk group had greater enrichment of CD8 + T cells, cytolytic activity, HLA molecules, neutrophils, T cell co-stimulation, helper T cells, and tumor-infiltrating lymphocytes (TILs) (p < 0.05). The high-risk group presented increased APC co-inhibition, parainflammation, and type I interferon response (p < 0.05) (Fig. 5d). CIBERSORT analysis found the high-risk group had more M1 macrophages and T cells CD4 memory resting, while the low-risk group had more T cells CD4 memory activated and T cells CD8 (Fig. 5c, e). Most immune checkpoint genes—including CD44, CD276, LGALS9, HHLA2, TNFSF9, PDL1 and TNFSF15—were significantly upregulated in the high-risk group (Fig. 5f, g). TIDE analysis revealed that the high-risk group had significantly greater T cell exclusion and microsatellite instability (MSI) than the low-risk group did, indicating an immunosuppressive environment (Fig. 5h–k). However, Immunophenoscore (IPS) comparisons across the CTLA-4/PD-1 subgroups revealed no significant differences between the risk groups (Fig. S3a). Similarly, EaSIeR-predicted ICB response scores did not significantly differ by risk group (Fig. S3b), suggesting the limited predictive value of the risk score for immunotherapy response.

Fig. 5.

Assessment of immune status between high-risk and low-risk groups. a Distribution changes of immune-related cells between the two risk groups. b Boxplots of ESTIMATE, immune, stromal scores and tumor purity in high-risk and low-risk score groups. c Distribution patterns of immune cells in high-risk and low-risk groups. d Boxplots of 29 immune cells and immune functions between high-risk and low-risk groups. e Boxplots showing infiltration patterns of 22 immune cell subpopulations between high-risk and low-risk groups. f Boxplots depicting expression levels of immune checkpoints in high-risk and low-risk groups. g Correlations among 13 immune checkpoints. Comparison between high-risk and low-risk score groups for: TIDE scores (h), T cell dysfunction (i), T cell exclusion (j) and microsatellite instability scores (k)

Tumor mutation burden

Genomic mutation analysis between the high- and low-risk groups revealed distinct mutation patterns (Fig. 6a, b). The most frequently mutated genes in the high-risk group were KRAS (78%), TP53 (71%), CDKN2A (26%), SMAD4 (22%) and TTN (11%), whereas those in the low-risk group they were TP53 (48%), KRAS (45%), SMAD4 (21%), TTN (14%) and CDKN2A (10%). Fisher’s exact test revealed significant differences in mutation frequency between groups (Fig. 6c, f, g). Co-occurrence and exclusivity analyses identified PREX1-KMT2C as the top comutated pair and COL6A2-TP53 as mutually exclusive in the high-risk group; DST-TP53 was the strongest cooccurring pair in the low-risk group (Fig. 6d). CNV events, including amplifications and deletions, were more common in high-risk patients (Fig. 6e), suggesting increased genomic instability in the high-risk subgroup, which may contribute to aggressive tumor behavior. However, Kaplan–Meier curves revealed no difference in survival between the high and low TMB groups, which may be attributed to the generally poor prognosis of PAAD patients (Fig. S4a). The wilcoxon test revealed a significantly greater TMB in the high-risk group than in the low-risk group, and the TMB was positively correlated with the risk score (Fig. 6h). Patients with high TMB and low risk had better survival than those with high TMB and high risk (Fig. 6i). There were no significant differences in the Mutant-Allele Tumor Heterogeneity (MATH) scores between the risk groups, (Fig. S4b), but the Microsatellite Stable (MSS) and Microsatellite Instability (MSI) distributions differed significantly between the groups (Fig. S4c). GSVA (Gene Set Variation Analysis) revealed six hallmark pathways significantly enriched in the low-risk group, including myogenesis, KRAS signaling down, pancreas beta cells, bile acid metabolism, angiogenesis, and hedgehog signaling (Fig. S4d).

Fig. 6.

Mutation analysis. a Top 20 frequently mutated genes in high-risk and (b) low-risk groups. c Heatmap showing significantly different mutations between high-risk and low-risk groups. d Heatmap displaying co-occurrence or mutual exclusivity of 20 mutated genes between high and low-risk groups. Green and brown colors represent the probability of two genes experiencing co-occurring or mutually exclusive mutations, with deeper colors indicating higher probability. e Changes in gene copy number levels between high-risk and low-risk patients, where red indicates significantly increased copy numbers and blue indicates significant copy number deletions. f Bar plot showing significantly different mutations between high-risk and low-risk groups. g Forest plot displaying the top 12 significantly mutated genes in the high-risk group compared to the low-risk group. *p < 0.05, **p < 0.01. h Correlation between TMB and risk score; boxplot showing TMB differences between high and low-risk groups. i Kaplan-Meier survival curves for high and low groups in the combined TMB-risk model

Gene expression verification and distribution analysis

RPN1 and PML were identified as key disulfidptosis-related genes on the basis of their prognostic and expression profiles (Fig. 7a), both of which are elevated in pancreatic cancer and associated with poor prognosis. This overexpression was validated using TCGA, GTEx, GSE28735, and UALCAN datasets (Fig. 7b, c). Correlation analysis revealed a significant positive association between RPN1 and PML in both TCGA and GSE28735 cohorts (Fig. 7d). Functional enrichment analysis revealed that RPN1 was involved in cell growth and death (e.g., p53 signaling pathway, oocyte meiosis) and immune regulation pathways (e.g., hematopoietic cell lineage, complement and coagulation cascades) (Table S4). GSEA linked RPN1 to the G2/M checkpoint, glycolysis, and hypoxia (Fig. S5a, S5c, S5e), suggesting roles in proliferation and metabolic reprogramming. Previous studies further reported that RPN1 is a crucial subunit of the oligosaccharyltransferase complex, playing roles in N-linked glycosylation and endoplasmic reticulum stress response, and may contribute to tumor immune evasion [47]. PML was enriched in apical plasma membrane, hormone activity, PI3K-Akt signaling pathway, and phenylalanine metabolism (Table S5). GSEA revealed that PML was positively correlated with glycolysis and negatively correlated with pancreas beta cells (Fig. S5b, S5d, S5f). These findings are consistent with earlier reports that PML regulates DNA damage response, apoptosis, senescence, and tumor suppression [48]. HPA confirmed the protein expression trends: PML was elevated in tumors, while RPN1 was increased in normal tissues. Subcellular localization revealed RPN1 in the endoplasmic reticulum and cytoplasm, and PML in the nucleoplasm (Fig. 7e, f). Taken together, both RPN1 and PML may contribute to tumor progression through multiple biological processes beyond disulfidptosis.

Fig. 7.

Gene expression verification and distribution analysis. a Kaplan-Meier survival curves for RPN1 and PML. b Relative expression levels of RPN1 and PML in 178 pancreatic cancer tissues and 171 normal tissues from TCGA merged with GTEx database (left). Relative expression levels of RPN1 and PML in the GSE28735 dataset (N = 42, T = 42) (right). c Protein expression validation based on the UALCAN database. d Correlation analysis between RPN1 and PML expression. e Immunohistochemical images of RPN1(left) and PML(right) in pancreatic normal and tumor tissues. f Distribution of RPN1 (left) and PML(right) expression in different subcellular structures

Immune infiltration analysis of RPN1 and PML

The risk score was positively correlated with parainflammation and APC co-inhibition. Both RPN1 and PML were also positively associated with γδ T cells, indicating potential associations with nonconventional T cell mediated immunity (Fig. 8a). ssGSEA and CIBERSORT analyses suggested that in the high-risk group, high RPN1 expression may be linked to APC co inhibition/stimulation, parainflammation, Type I IFN Reponse and Macrophages M0, similar trends were observed for PML (Fig. 8b, c). Multiple deconvolution algorithms suggested that RPN1 may be negatively correlated with ImmuneScore, while the risk score may be positively associated with TumorPurity (Fig. S6a). RPN1 was also potentially associated with T cell CD4 + Th2, NK cell and Macrophage M1; PML may be linked to Macrophage M1 and Neutrophil (Fig. S6b, S6c). Overall, the risk score appeared to correlate with immune cell infiltration and checkpoint expression (Fig. 8d; Fig. S6d), GSVA based on hallmark gene sets suggested associations between RPN1, PML, the risk score, and pathways such as UV response up, TGF-β, P53, and mTORC1 signaling, highlighting potential associations with tumor-related pathways that warrant further investigation (Fig. S6e), which was supported by TIMER2.0 data (Fig. S6f).

Fig. 8.

Comprehensive analysis of immune-related correlations. a Correlation matrix showing the relationships between RPN1, PML, the risk score and components of the cancer immune cycle and metabolic pathways. b Boxplot visualization of immune status scores for RPN1, PML and risk score expression levels, where higher scores indicate enhanced immune activity. The horizontal axis represents immune gene sets and the vertical axis represents immune scores. c Boxplot analysis of infiltration patterns across 22 distinct immune cell subpopulations. d Correlation analysis between of RPN1, PML, the risk score and immune checkpoint molecules

Drug sensitivity

Chemotherapy is central to pancreatic cancer treatment, especially for advanced cases, although patient responses vary. Using drug data (Table S6), the key agents used included gemcitabine, oxaliplatin, irinotecan, abraxane, and cisplatin. Drug sensitivity prediction using the R package oncoPredict suggested that the high-risk and low-risk groups may exhibit differential responses to certain chemotherapeutic agents (Table S7). For example, drugs such as acetalax, afatinib, osimertinib, and trametinib show lower IC50 values in the high-risk group (Fig. 9a). RPN1 expression showed potential correlations with sensitivity to venetoclax, sorafenib and oxaliplatin, while PML may be linked to vorinostat, sorafenib, and tamoxifen responses (Fig. S7a, S7b). Both genes were associated with higher epithelial–mesenchymal transition (EMT) scores, suggesting possible involvement in metastasis-related processes (Fig. S7c, S7d). The correlation between gene expression and the drug IC50 is shown in Fig. S7e–g. CellMiner pharmacogenomic analysis indicated potential links between RPN1, PML, and sensitivity to 51 anticancer drugs (Table S8). RPN1 was positively correlated with nitazoxanide and abiraterone sensitivity but negatively with neratinib and erlotinib; PML was positively correlated with clofarabine and gemcitabine (Fig. 9b, c), indicating potential associations with drug sensitivity. Overall, these findings provide preliminary insights into subgroup-specific drug sensitivity and warrant further validation in experimental studies.

Fig. 9.

Drug sensitivity analysis. a Drug sensitivity analysis between the high-risk and low-risk groups. b Correlation between drug sensitivity and the expression of RPN1. c Correlation between drug sensitivity and the expression of PML

Expression patterns of RPN1 and PML in the tumor microenvironment

To explore the roles of RPN1 and PML in immune infiltration, we performed scRNA-seq analysis using the GSE212966 dataset. The cell type composition of the pancreatic tumor microenvironment was mapped (Fig. 10a; Fig. S8a, S8b). Consistent with the findings of bulk RNA-seq, PML expression was lower in tumors than in normal tissues (Fig. 10b). 3D clustering identified eight majorcell types (Fig. 10c). Cell type–specific analysis revealed that RPN1 was mainly expressed mainly in macrophages, with low expression in T cells, NK cells, and B cells. In contrast, PML was selectively expressed in endothelial cells (Fig. 10d–h). Spatial and quantitative analyses confirmed the association of RPN1 with macrophages (Fig. 10i, j), suggesting its involvement in creating a macrophage-rich immunosuppressive environment in pancreatic cancer.

Fig. 10.

Validation of key genes in the single-cell dataset GSE212966. a Dimensionality reduction clustering plot of immune microenvironment in normal tissues and tumor tissues. b Violin plot showing differential expression of RPN1 and PML in normal and tumor samples from single-cell dataset. c Three-dimensional plot of 13 identified cell types. d Three-dimensional plot displaying expression levels of RPN1 across 13 identified cell types. e Three-dimensional plot displaying expression levels of PML across 13 identified cell types. f Bubble plot showing expression levels of RPN1 and PML in the TME. g Expression pattern of RPN1 in the tumor microenvironment. h Expression pattern of PML in the tumor microenvironment. i Violin plot depicting RPN1 expression levels in various subgroups of adjacent nontumor and tumor samples. j Violin plot depicting PML expression levels in various subgroups of adjacent nontumor and tumor samples

Pseudotime dynamics of RPN1 across macrophage subsets and cell–cell communication profiling of RPN1-high and -low macrophages

To further explore the heterogeneity of macrophages on the basis of RPN1 expression, we stratified the macrophages into RPN1-high and RPN1-low groups and performed trajectory inference via the Slingshot algorithm to evaluate transcriptional diversity and lineage dynamics (Fig. 11a; Fig. S8c). The distribution of RPN1 expression across distinct macrophage cell states is illustrated in Fig. 11b. Subsequent cell–cell communication analysis quantified the number and intensity of interactions between RPN1-high and RPN1-low macrophages, as well as their interactions with other immune and stromal cell types (Fig. 11c–e). As shown in the heatmap (Fig. 11f), RPN1-high macrophages presented a greater overall communication probability, with notably increased activity in the SPP1 signaling pathway. Given that pancreatic cancer progression is closely linked to immune evasion, microenvironmental remodeling and EMT, we further examined the intercellular communication landscape of the TGF-β signaling axis specifically among immune cell populations within tumor tissues (Fig. S8d). Figure 11g, h illustrate the ligand–receptor interactions between RPN1-labeled macrophages and other cell types. Notably, RPN1-high macrophages establish communication through critical ligand–receptor pairs such as MIF–(CD74 + CXCR4) and SPP1–CD44, indicating potential involvement in immune modulation and stromal interactions. Metabolic pathway analysis revealed that RPN1-high macrophages were significantly enriched in a range of metabolic processes compared with RPN1-low macrophages, suggesting a distinct metabolic phenotype potentially relevant to their immunosuppressive or protumor functions (Fig. 11i).

Fig. 11.

Pseudotime and cell-cell interaction analysis. a Pseudotime trajectory analysis of macrophages. b Expression level of RPN1 in relation to pseudotime and changes in cell states. c–e Cell communication analysis showing the number and strength of interactions between RPN1-high macrophages and other cell types. f Heatmap visualization showing that RPN1-high macrophages exhibit a greater probability of communication. g–h Ligand–receptor interactions between different cell types and RPN1-marked macrophages analyzed by cell communication analysis. i Pathway enrichment analysis

Discussion

Pancreatic adenocarcinoma (PAAD) is recognized as one of the most lethal solid tumors, with therapeutic resistance posing a major limitation to the efficacy of both conventional chemotherapy and immunotherapy. In recent years, studies of cell death mechanisms have uncovered multiple novel forms of programmed cell death. Among these, disulfidptosis, a newly described death mode triggered by disulfide accumulation leading to F-actin cytoskeleton collapse, has demonstrated unique biological significance in PAAD [49].

Disulfidptosis was first identified in glucose-deprived cells with high SLC7A11 expression. SLC7A11, a core component of the system Xc⁻ transporter, facilitates cystine uptake and glutathione (GSH) synthesis, thereby enhancing antioxidant capacity in tumor cells. In PAAD, SLC7A11 is frequently upregulated and the tumor’s heavy reliance on glucose metabolism renders it more vulnerable to disulfidptosis under nutrient deprivation [50]. Studies have shown that disulfidptosis is independent of lipid peroxidation and instead involves F-actin depolymerization induced by excessive disulfide bond formation, ultimately leading to cell contraction and death.Key regulatory pathways implicated in this process include Rac1 activity, the β-catenin/MYC axis and GLUT1-mediated glucose transport [51]. MYC, a commonly activated factor in PDAC, upregulates GLUT1, promoting glucose uptake and thereby increasing susceptibility to disulfidptosis—offering a precise metabolic target for therapy. Recent studies have also developed prognosticmodels based on disulfidptosis-related gene (DRG) expression profiles, linking the pathway to immune phenotypes, PD-L1 expression and T cell infiltration in PDAC [52]. These findings support the rationale for integrating disulfidptosis with immunotherapy strategies. Moreover, GLUT1 inhibitors or glucose deprivation have been shown to effectively induce disulfidptosis in vitro, particularly in SLC7A11-high PDAC cells [53]. Collectively, disulfidptosis representsa unique metabolic stress-induced death mechanism that may explain treatment resistance in PAAD and serve as a core element in metabolic targeting and personalized therapeutic strategies. Emerging approaches, such as circular RNA-based vaccines, may further complement these strategies in precision oncology [54]. Similar epigenetic vulnerabilities, such as methylation of FAM110C leading to synthetic lethality with ATR/CHK1 inhibitors, highlight additional opportunities for targeting DNA damage repair pathways in PDAC, potentially synergizing with disulfidptosis-related metabolic stresses [55].

Machine learning algorithms provide powerful tools for analyzing multi-omics data. In this study, we developed a DRG-based prognostic model for PAAD using TCGA data and validated it in the GSE28735 dataset. Recent advancements in multi-omics integration combined with machine learning have further enabled the identification of molecular subtypes in pancreatic cancer, such as basal-like subtypes, and the construction of robust prognostic signatures using algorithms like ridge regression, which demonstrate correlations with drug sensitivity and pathways like KRAS/MAPK [56]. Univariate Cox regression identified 19 prognostically relevant DRGs, including NDUFB11, NDUFA11, MYL6B, INF2, MYH14, TRIP6, ACTN1, PDLIM1, ANP32E, MYH9, CHCHD3, ACTN4, IQGAP1, PML, NLN, CD2AP, IPO7, RPN1 and GLUD1. We applied 117 combinations of algorithms to both training and validation cohorts to identify the optimal model and mitigate bias from algorithm selection. The optimal model—StepCox[both] + RSF—included 8 genes: NDUFA11, MYH14, TRIP6, ACTN1, ANP32E, PML, CD2AP and RPN1, which were used to estimate OS in PAAD patients. Based on median risk scores, patients were stratified into high- and low-risk groups. Survival analysis revealed a significant difference in survival between the two groups. ROC curve analysis confirmed the predictive accuracy and robustness of the model. We also constructed a nomogram integrating risk score and clinical features to quantitatively estimate 0.5-, 1-, 2- and 3-year survival rates for individual PAAD patients, aiding in clinical decision-making.

While StepCox[both] + RSF model demonstrated excellent performance in the training set (C-index 0.9), external validation in the GSE28735 cohort yielded a moderate C-index of 0.6, indicating potential overfitting. This performance gap may arise from differences in cohort characteristics, such as sample size (TCGA: n = 178; GSE28735: n = 90), platform variations, as well as the high dimensionality of gene features relative to sample numbers, which can lead to model instability in machine learning applications. Consequently, while the model identifies promising DRG signatures and associations with immune infiltration and drug sensitivity, these findings should be interpreted as preliminary and hypothesis-generating. Prospective clinical studies and further external validations are essential to confirm the model’s generalizability and the reliability of downstream inferences, such as the roles of RPN1 and PML in TME dynamics.

Further analysis of immune infiltration, risk stratification and genomic features revealed complex interactions between the TME and patient prognosis. Heatmaps showed higher immune cell expression in the high-risk group, suggesting an active but potentially dysfunctional immune response. However, ESTIMATE analysis showed no significant differences in immune, stromal, or tumor purity scores between risk groups, suggesting that immune cell function and interaction—rather than mere abundance—may be more critical in determining prognosis.ssGSEA revealed that the low-risk group had significantly higher enrichment scores for CD8 + T cells, cytolytic activity, HLA, neutrophils, T cell co-stimulation, helper T cells and TILs (p < 0.05), indicative of a stronger anti-tumor immune response. In contrast, the high-risk group showed elevated scores for APC co-inhibition, parainflammation and type I IFN response, indicating an immunosuppressive TME potentially contributing to poor prognosis. CIBERSORT analysis supported these findings, revealing that M1 macrophages and resting CD4 + memory T cells were more abundant in the high-risk group, while activated CD4 + memory T cells and CD8 + T cells were enriched in the low-risk group, highlighting the protective role of immune activation in PAAD. Immune checkpoint analysis revealed upregulation of CD44, CD276, LGALS9, HHLA2, TNFSF9, PDL1 and TNFSF15 in the high-risk group, suggesting potential therapeutic targets. TIDE analysis showed significantly higher T cell exclusion and MSI scores in the high-risk group, supporting the hypothesis that immune escape mechanisms may contribute to resistance to immunotherapy. The clinical application of immune checkpoint inhibitors in PC must consider associated endocrinopathies, as highlighted in recent reviews of their side effects [57]. However, both IPS and EaSIeR analyses indicated no significant difference in predicted response to immunotherapy between risk groups, suggesting that risk score alone may not predict PDAC immunotherapy response, potentially due to its intrinsic low immunogenicity. Genomic analyses revealed distinct mutational patterns: in the high-risk group, KRAS (77%) and TP53 (70%) were most frequently mutated, while in the low-risk group, TP53 (48%) and KRAS (45%) were dominant. Fisher’s exact test confirmed significant differences in mutation frequencies. CNV amplifications and deletions were more frequent in the high-risk group, indicating greater genomic instability as a driver of aggressiveness.

TMB was positively correlated with risk score, and patients in the high TMB + low-risk subgroup had significantly better outcomes than those in the high TMB + high-risk subgroup, highlighting the potential of TMB–risk score integration for personalized treatment. However, no significant difference in MATH scores or OS was observed between high and low TMB groups, likely reflecting the generally poor prognosis of PAAD. GSVA revealed that hallmark pathways such as myogenesis, KRAS signaling down and angiogenesis were enriched in the low-risk group, potentially contributing to a less aggressive phenotype. These findings suggest that integrating immune, genomic and pathway-level data may enhance the predictive power of risk models and inform therapeutic strategies for PAAD.

From the prognostic and expression analysis, RPN1 and PML were identified as core genes in PAAD, with high expression levels associated with poor prognosis. Data from TCGA, GTEx and GSE28735 consistently showed significant upregulation of RPN1 and PML in tumor tissues. Their positive correlation suggests a possible synergistic role in tumor progression.

Functional enrichment analyses (GO, KEGG, GSEA) revealed that RPN1 primarily regulates cell cycle, p53 signaling and metabolic reprogramming, while PML is involved in extracellular matrix remodeling, PI3K-Akt signaling and glycolysis, implicating both genes in tumor proliferation, metabolism and microenvironment regulation.

Risk score was positively correlated with tumor purity and negatively with immune score, suggesting that high-risk patients exhibit a more immunosuppressive TME. Single-cell RNA sequencing further revealed that RPN1 was predominantly expressed in macrophages, while PML was enriched in endothelial cells. HALLMARK pathway analysis indicated strong associations of RPN1, PML and risk score with TGF-β, mTORC1 and other pro-tumorigenic signaling pathways, reinforcing their roles in tumor invasion and immune modulation. However, this study is limited by its retrospective design and reliance on public datasets, and prospective clinical validation is warranted.

In conclusion, RPN1 and PML are not only prognostic biomarkers in PAAD but may also actively contribute to tumor progression by regulating immune and metabolic pathways. They represent promising targets for the development of novel therapeutic strategies, including metabolic and immunomodulatory interventions. Future studies should explore the functional roles of RPN1 and PML in vivo and assess their potential as therapeutic targets in preclinical models.

Conclusion

In summary, our study delineates the complex interplay among disulfidptosis, biological function, immune infiltration and clinical outcomes in pancreatic cancer. The identified prognostic model and potential immunotherapeutic targets offer promising avenues for advancing precision medicine strategies in the management of PAAD and warrant further validation in prospective clinical studies.

Author contributions

Y.P. contributed to the writing, reviewing and editing of the manuscript and was responsible for figure preparation. M.J.C. and T.Y. searched public databases and collected the data. W.J.M., X.H.S., J.Y.Z., J.J.L. and W.K.L. performed the data analysis. Y.T.Z. and H.W.H. contributed to the data analysis. Y.C.Q. was responsible for proofreading the manuscript. Y.E.L. provided critical insights, revised the manuscript for important intellectual content. Y.C.W. provided the research concept, supervised the study, reviewed the manuscript and offered technical support. All authors have read and approved the final version of the manuscript.

Funding

This research was supported by the Liaoning Provincial Natural Science Foundation Joint Fund (Doctoral Research Start-up Project) (2023-BSBA-291).

Data availability

The public datasets analyzed in this study are available from The Cancer Genome Atlas (TCGA) portal ([https://portal.gdc.cancer.gov/] (https:/portal.gdc.cancer.gov)) and the Gene Expression Omnibus (GEO) database ([http://www.ncbi.nlm.nih.gov/geo/] (http://www.ncbi.nlm.nih.gov/geo)). Further information is available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The TCGA and GEO databases are publicly available. Users can download the data for free for research purposes and publish related articles. The authors declare no ethical issues or conflicts of interest.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.