Metabolism pathway-based subtyping in pancreatic adenocarcinoma: an integrated study by bulk RNA-sequence and machine learning algorithms

Abstract

Background:

Pancreatic adenocarcinoma (PAAD) is highly aggressive, and its tumor microenvironment has significant metabolic and immune microenvironment complexity and genomic instability. In this study, by integrating the metabolic pathway activity score and clinical data, we constructed a novel risk assessment model to reveal the unique biological behavior and clinical significance behind different PAAD subtypes.

Methods:

In this study, the transcriptome and clinical data of TCGA and GSE57495 databases were integrated to explore the interaction between metabolic pathways. Based on unsupervised clustering analysis of pathway activity and survival prognosis, patients with PAAD were classified into metabolic subtypes with significant prognostic differences. Subsequently, we assessed the heterogeneity of these subtypes in terms of clinical outcomes, genomic characteristics, and immune microenvironment composition. Based on the differentially expressed genes (DEGs) among metabolic subtypes, a clinical prognostic risk model and nomogram were constructed, which were double-validated by GSE57495-independent cohort and GSE57495 + TCGA-PAAD combined cohort. Finally, the correlations between risk scores (RSs) and signaling pathway activity and tumor immune microenvironment characteristics were evaluated.

Results:

Based on metabolic pathway correlation and prognostic information, 240 patients in the TCGA-PAAD and GSE57495 datasets were divided into three subgroups. There were significant differences between subgroups in gene expression, pathway activity, clinical prognosis, and immune infiltration characteristics among the subtypes. Using machine learning algorithms, an RS model was constructed from DEGs among the subgroups, with the random forest method showing the best performance. A nomogram integrating the RS and clinical indicators demonstrated excellent predictive accuracy for 1-, 3-, and 5-year survival rates, confirming the RS as an independent prognostic factor. High- and low-risk groups exhibited significant differences in immune infiltration, pathway activity, and gene mutations. Drug sensitivity analysis showed that the high-risk group was more sensitive to AZD6244, ABT737, and other drugs.

Conclusion:

This study stratified patients with PAAD into three subgroups based on metabolic pathways and prognostic information, revealing significant differences in clinical outcomes, immune characteristics, and genetic mutations. The robust RS model developed from these findings demonstrated strong predictive power for patient survival and identified promising therapeutic strategies, providing valuable insights for advancing precision medicine in PAAD.

Introduction

Pancreatic adenocarcinoma (PAAD) is one of the most aggressive malignant tumors with poor prognosis. Due to its insidious onset, rapid progression, and high resistance to conventional treatment, the clinical management of this disease faces severe challenges ^[1,2^]. Epidemiological data show that there are about 495 000 new cases and 466 000 deaths each year globally, and the mortality rate is almost equal to the incidence^[3^]. The lack of efficient early diagnostic biomarkers and persistent treatment resistance are still the key factors hindering the improvement of clinical interventions^[4^]. In response to these challenges, current studies are concentrating on the development of more effective treatment strategies^[5–7^] and accurate prognostic tools^[8–10^]. In recent years, the progress of immunotherapy and targeted therapy has provided new opportunities to improve the survival and prognosis of patients, particularly in modulating the tumor microenvironment and targeting KRAS mutations^[11–13^]. Clarifying the molecular mechanisms of PAAD, including tumorigenic pathways, immune escape strategies, and drug resistance mechanisms, is crucial for overcoming current therapeutic limitations.

Metabolic reprogramming is a hallmark of cancer, playing a pivotal role in tumor growth, survival, and adaptive regulation^[14^]. In PAAD, enhanced glycolysis, disrupted lipid metabolism, and increased amino acid utilization are key pathways that enable tumor cells to fulfill their energy and biosynthetic demands during rapid proliferation^[15^]. These metabolic changes not only provide essential substrates for cells but also activate multiple signaling pathways and promote the accumulation of metabolic intermediates, thereby accelerating tumor progression, promoting immune escape, and enhancing tumor adaptability to the microenvironment^[16^]. In addition, abnormal metabolic activity is also closely related to treatment resistance^[17^]. For instance, glycolytic by-products can alter the tumor microenvironment by suppressing immune cell function, thereby reducing the efficacy of immunotherapy^[18,19^]. Similarly, aberrant lipid metabolism supports membrane biogenesis and energy storage, facilitating tumor resilience under therapeutic stress. Despite our understanding of tumor metabolism, it remains a great challenge to integrate these complex metabolic pathway features into clinically usable prognostic assessment models^[20^]. Current studies often focus on single pathway, and lack a systematic understanding of the overall and dynamic changes of tumor metabolic network^[21^].

Recent developments in machine learning have opened new avenues for integrating complex, high-dimensional multi-omics datasets, which has facilitated the development of prognostic models with improved robustness^[22,23^]. In cancer research, various algorithms, such as random forests, LASSO, and Cox regression, have been widely used in a variety of cancer studies to identify key biomarkers, stratify patients by risk, and predict survival outcomes^[24,25^].

In this study, we sought to classify patients with PAAD into distinct subgroups based on metabolic pathway activity and survival prognosis. Through the integration of genomic profiles and clinical characteristics, we constructed a robust risk scoring system. This framework enables stratification of patient prognosis and highlights potential therapeutic targets, thereby contributing to the advancement of precision medicine in PAAD. In alignment with ethical publishing standards, this study adheres to the TITAN 2025 guidelines on AI transparency in scientific writing^[26^].

HIGHLIGHTS

• For the first time, based on metabolic pathway activities, PAAD patients were classified into three subtypes (Cluster A/B/C), revealing distinct differences in clinical outcomes, immune microenvironments, and gene mutations.

• Developed a random forest-based risk scoring model with significant predictive efficacy for 1-, 3-, and 5-year survival rates (AUC > 0.9), enabling precise prognosis assessment for patients with PAAD.

• Identified high-risk patients with PAAD having stronger tumor stemness and higher mutation burden, and found them more sensitive to targeted drugs like AZD6244, providing a basis for personalized treatment strategies.

Materials and methods

Data collection and preprocessing

The GSE57495 dataset, containing survival information for 63 PAAD samples, was obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). Additionally, transcriptomic and clinical data for 177 PAAD samples were retrieved from the TCGA-PAAD cohort via the UCSC Xena platform (https://xena.ucsc.edu/). Batch effects between datasets were corrected using the “limma” and “sva” packages in R (version 4.2.2), ensuring data comparability and consistency.

Correlation and survival analysis of metabolic pathways

Metabolic pathway annotations were sourced from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database in the MsigDB repository, focusing on 41 metabolism-related pathways. Gene Set Variation Analysis (GSVA) was performed to calculate pathway activity scores. Correlation analyses were conducted to explore interactions among metabolic pathways. Survival analysis was carried out using the “survival” and “survminer” packages in R, with Kaplan-Meier (KM) analysis identifying pathways significantly associated with overall survival (OS). Pathways with a P-value <0.05 were considered statistically significant.

Unsupervised clustering of metabolic subtypes

Unsupervised clustering of PAAD samples was performed based on the GSVA scores of significantly associated metabolic pathways using the ConsensusClusterPlus algorithm. The optimal number of clusters was determined by evaluating the cumulative distribution function and delta area plot. Heatmaps were generated using the pheatmap package to visualize the relationships between clinical characteristics, gene expression patterns, and metabolic subtypes.

Differential gene expression and enrichment analysis between subgroups

Differential gene expression analysis was conducted between metabolic subtypes using the “limma” package. Genes with an adjusted P-value < 0.05 and |logFC| > 1 were considered differentially expressed. Functional enrichment analyses, including Gene Ontology (GO), KEGG, and Reactome pathway analyses, were performed using the clusterProfiler package in R to investigate the biological processes and pathways enriched in these genes.

Mutation analysis

The mutation landscape of PAAD samples was characterized using mutation data from TCGA and GSE datasets. The maftools package in R was utilized to visualize mutation frequency and mutation types, including single-nucleotide variations and insertions/deletions (indels). The mutation spectrum was analyzed for its potential associations with metabolic subtypes and clinical outcomes.

Immune infiltration analysis

Single-sample Gene Set Enrichment Analysis (ssGSEA) was used to estimate the abundance of 28 immune cell types in PAAD samples. Immune infiltration levels were compared across different metabolic subtypes to identify variations in immune cell populations. Spearman correlation analysis was performed to evaluate the relationships between immune cell infiltration levels and metabolic pathway activities. Additionally, the expression of immune checkpoints, including PD-1, PD-L1, and CTLA4, was analyzed.

Machine learning algorithms for metabolic prognostic model construction

Prognostic genes were identified by univariate Cox regression analysis from the TCGA-PAAD and GSE57495 cohorts. Expression data were Z-score normalized, with the TCGA cohort used for training and GSE57495 plus the combined dataset for validation. Ten survival modeling algorithms – random survival forest (RSF), LASSO Cox, GBM, Survival-SVM, SuperPC, ridge regression, PLSRcox, CoxBoost, stepwise Cox, and elastic net – were applied individually or in pairwise combinations (117 strategies in total). Optimal parameters were determined by cross-validation. Model performance was evaluated using Harrell’s concordance index (C-index), and patient risk scores (RSs) were calculated to stratify individuals into high- and low-risk groups based on the median score. All analyses were performed in R using standard survival modeling packages.

Risk model construction and evaluation

Kaplan-Meier (KM) survival analysis, receiver operating characteristic (ROC) curves, and area under the curve (AUC) metrics were used to evaluate the risk model’s predictive performance for 1-, 3-, and 5-year survival rates. Stratified analysis of clinical characteristics was conducted for high- and low-risk groups. External validation was conducted using the GSE57495 dataset and a combined cohort of the GSE57495 dataset and the TCGA-PAAD cohort, with RSs calculated using the same methodology.

Construction of a nomogram

A nomogram was developed by integrating the RS with clinical factors such as age, sex, and tumor-node-metastasis (TNM) stage. Multivariate Cox regression analysis was used to identify independent prognostic factors. The nomogram’s calibration was assessed using calibration plots, and its discrimination ability was evaluated with ROC curves and decision curve analysis. Validation was conducted using external datasets to ensure robustness.

Correlation analysis of RS with immune infiltration

The relationship between RSs and immune cell infiltration levels was explored using ssGSEA results. Stromal, immune, and ESTIMATE scores, along with tumor purity, were calculated using the ESTIMATE algorithm. Correlation analyses were conducted to determine associations between these parameters and RSs.

Prediction of drug susceptibility based on risk assessment

Drug sensitivity prediction was conducted using the oncoPredict R package. The half-maximal inhibitory concentration (IC₅₀), representing the drug concentration required to inhibit 50% of cellular activity, was estimated for various anticancer drugs based on transcriptional profiles of all samples. Patients were categorized into high-risk and low-risk groups based on their calculated RSs. Differences in IC₅₀ values between these groups were analyzed to identify variations in drug sensitivity.

Correlation analysis between RS and stemness

Stemness indices, including mDNAsi (based on DNA methylation) and mRNAsi (based on mRNA expression), were calculated using the one-class logistic regression algorithm to assess tumor cell stemlike characteristics. Patients were stratified into high- and low-risk groups based on RSs, and statistical comparisons of stemness indices between groups were performed using the Wilcoxon rank-sum test. Pearson correlation analysis was conducted to evaluate the relationship between RSs and stemness indices, with significance defined as P < 0.05.

Statistical analysis

The statistical analysis was completed using the R version (4.2.1). For binary categorical variables, the Wilcoxon rank-sum test was used to compare groups, and the Kruskal–Wallis test was used for multiclass comparisons. KM survival analysis was used for survival assessments. Univariate Cox regression analysis was used to identify prognosis-significant variables, whereas a multivariate Cox regression model was constructed to develop the predictive model. Variables with a P < 0.05 were deemed statistically significant.

Results

Metabolic pathways network in pancreatic cancer

To investigate interactions among metabolic pathways in pancreatic cancer, we utilized gene expression data from the TCGA-PAAD dataset. To include data from multiple sources to improve representation, we integrated data from the pancreatic cancer cohort GSE57495 data set in the GEO database to create a pooled data set of 17 247 genes with 240 samples (hereafter referred to as the pooled cohort). Principal component analysis (PCA) was used to evaluate differences between the two datasets. The pre-correction PCA (Fig. 1A) demonstrated variability, particularly along the first principal component (Dim1, 62.5%), indicating the presence of batch effects. To mitigate these effects, we performed batch effect correction using the “limma” and “sva” software packages in R. Post-correction PCA (Fig. 1B) showed that the distributions of the two datasets became more aligned, confirmed that the batch effect correction effectively reduced technical variability, and also explained the biological consistency of metabolic characteristics of patients with pancreatic cancer from different sources.

Figure 1.

Metabolic pathway correlation network in pancreatic cancer and subtyping of patients with pancreatic cancer based on metabolic pathways. (A) PCA based on the GSE62465 and TCGA datasets. (B) Overall PCA plot combining the datasets. (C) Network diagram of metabolic pathways associated with clinical outcomes. Red nodes represent metabolic pathways, while pink and green edges indicate positive and negative correlations with risk factors and protective factors, respectively. Node size reflects the P-value from the Cox proportional hazards model analysis. (D) Consensus clustering identified three distinct subtypes (Clusters A, B, and C) based on metabolic pathways. The consensus heatmap shows high intra-cluster consistency. (E) Box plots demonstrate the expression levels of 27 metabolic pathways across the three clusters. Statistical significance is indicated (*P < 0.05, **P < 0.01, ***P < 0.001).

Subsequently, 41 metabolic pathways were selected based on the KEGG database. By GSVA and KM survival analysis, 27 metabolic pathways were significantly associated with survival, including fatty acid metabolism, amino acid metabolism, nucleotide metabolism, and glucose metabolism. To explore the interaction of these metabolic pathways in pancreatic cancer, we performed correlation analysis of different pathways. The metabolic pathway network (Fig. 1C) clearly illustrates the interaction of these pathways. Notably, glutathione metabolism (ko00480, P = 0.0126), cysteine and methionine metabolism (ko00270, P = 0.018), and starch and sucrose metabolism (ko00500, P = 0.0049) play crucial roles within the network and are significantly associated with poor prognosis, consistent with previous reports. Additionally, although not statistically significant (P = 0.075), α-linolenic acid and linoleic acid metabolism (ko00592) appear to have a protective role. Survival analysis suggests that patients with higher α-linolenic and linoleic acid metabolism have better outcomes (P = 0.020). This may be related to its anti-inflammatory, anti-oxidation, and metabolic regulation effects.

To explore the metabolic heterogeneity of PAAD, we assessed the enrichment scores of 37 differentially expressed metabolic pathways using GSVA. Unsupervised cluster analysis using Consensual-ClusterPlus determined that using three subgroups was the best number to classify pancreatic cancer metabolic subgroups based on hierarchical clustering. Among the 240 patients with PAAD, 57 (23.75%) were divided into Group A, 111 (46.25%) into Group B, and 72 (30%) into Group C. PCA further confirmed clear differences in transcript levels between these isoforms (Fig. 1D).

We next examined the differential expression of 27 metabolic pathways among subtypes (Fig. 1E). Cluster A had a low overall expression level of most pathways, indicating that its metabolic activity was inhibited. However, some pathways reported to be protective against tumors (such as taurine and hypotaurine metabolism) were significantly upregulated in Cluster A compared to the other two subtypes, which is consistent with the results of previous survival analysis (Fig. 1C). Cluster B showed significantly high expression in the vast majority of metabolic pathways, reflecting robust metabolic activity and adaptability. Cluster C demonstrated “aggressive” metabolic characteristics, characterized by significantly elevated expression in energy metabolism pathways (e.g., glucose metabolism) and nucleic acid metabolism pathways (e.g., purine and pyrimidine metabolism), compared to Clusters A and B. We hypothesize that these activated pathways provide critical energy and nucleotides essential for tumor cell proliferation, thereby promoting tumor growth and invasion and contributing to the poor prognosis associated with Cluster C.

Survival characteristics, pathway activation, and mutational profiles differed among subgroups

Next, we analyzed the clinical and pathological characteristics of these subtypes. KM survival analysis revealed significant prognostic differences among the subtypes (Fig. 2A). Patients in Cluster C exhibited significantly poorer OS compared to those in Clusters A and B, as we hypothesized.

Figure 2.

Clinical features and pathway activation between three subgroups. (A) Kaplan–Meier survival curves indicate significant survival differences among the clusters (P = 0.008). Cluster A has the best prognosis, while Cluster C shows the worst survival. (B) Heatmap correlates clinical characteristics (Stage, T_stage, M_stage, N_stage, recurrence, etc.) with metabolic pathway expression levels. Yellow and blue represent high and low expression, respectively. (C–E) Pathway scores were calculated using GSVA for KEGG, HALLMARK, and REACTOME gene sets. Heatmaps illustrate the differential activity of pathways between the three subtypes (Clusters A–C). Yellow indicates high pathway activity, and blue indicates low pathway activity.

To integrate clinical features and pathway expression patterns, we generated a heatmap (Fig. 2B) to visualize the relationship between clustering results and clinical data. Among them, patients in Cluster A mainly belong to early-stage tumors to the early tumor stages (I–II), with low recurrence and metastasis rates, which is aligning with their protective metabolic characteristics and good prognosis. Conversely, Cluster C patients were primarily in advanced tumor stages (III–IV) and exhibited higher rates of recurrence and metastasis, reflecting their aggressive metabolic features. Additionally, Cluster C patients were generally older, further linking their metabolic profile to advanced disease stages and worse outcomes.

We further analyzed the distribution pattern of different signaling pathways in the three metabolic subtypes. In Cluster C, pathways associated with cell cycle regulation and mitosis including cell cycle checkpoints, E2F target genes, and metabolic precursors that promote rapid tumor growth were activated. In addition, pathways related to glucose metabolism and purine/pyrimidine metabolism were significantly activated in Cluster C and the PI3K/Akt/mTOR signaling pathway was significantly enriched (Fig. 2C, 2D, and 2E), indicating that this subtype is in a highly proliferative state and obtains energy through activation of these pathways. In contrast, Cluster A showed significant enrichment of immune response-related pathways, including interferon-α and interferon-γ signaling pathways. We hypothesized that the activation of these pathways may inhibit tumor invasion by enhancing the immune response, consistent with the favorable prognosis of Cluster A patients. Cluster B shows an intermediate state with activation of both invasive pathways (such as cell cyclin-related pathways) and protective pathways (such as immune-related pathways) and represents an intermediate-risk population.

One of the important causes of pancreatic cancer is mutations in genes, particularly KRAS and TP53, so we analyzed the rates of mutation in oncogenic pathways among the three subtypes (Supplemental Digital Content Figures S1A–C, available at: http://links.lww.com/JS9/H52). Cluster A had the lowest mutation rate (53.33%), with KRAS (37%) and TP53 (30%) being the predominant mutations, representing a standard tumor type with relatively low mutational burden (Supplemental Digital Content Figure S1A, available at: http://links.lww.com/JS9/H52). Cluster B showed an intermediate mutation rate (75.32%), primarily involving KRAS (55%) and TP53 (39%), along with significant increases in SMAD4 (14%) and CDKN2A (8%) mutations, suggesting partial genomic instability (Supplemental Digital Content Figure S1B, available at: http://links.lww.com/JS9/H52). Cluster C had the highest mutation rate (88.1%). In addition to KRAS and TP53, CDKN2A (17%), MUC16 (10%), and TPO (12%) had higher mutation rates. Previous studies have confirmed that CDKN2A mutations may enhance rapid tumor cell proliferation, while MUC16 and TPO mutations likely contribute to immune evasion and metabolic reprogramming (Supplemental Digital Content Figure S1C, available at: http://links.lww.com/JS9/H52). Missense mutation was the most common mutation type in the three metabolic subtypes. However, cluster C showed a higher proportion of multiple-hit mutations, which is indicative of the genomic complexity of aggressive tumors. In conclusion, our study demonstrated that three different metabolic subtypes of pancreatic cancer have different molecular pathway expression profiles and different gene mutation profiles, which are closely related to their prognostic characteristics.

Different immune microenvironment infiltration and gene expression between subgroups

The metabolic heterogeneity of pancreatic cancer is closely linked to changes in the tumor microenvironment. Various metabolic changes play a crucial role in tumor progression by promoting immune escape or impairing immune surveillance. To investigate the differences in immune infiltration among the metabolism-based subgroups, we used eight methods – MCPcounter, EPIC, CIBERSORT, xCell, quanTIseq, IPS, ESTIMATE, and TIMER (Fig. 3A) – to analyze the differences in immune cells among the three subgroups. PCA revealed significant differences in the distribution of immune cells between subtypes (Fig. 3B). Based on ssGSEA analysis, we observed that Cluster C exhibited higher immune infiltration levels, with significantly increases in T cell and natural killer (NK) cell activity (Fig. 3C). However, this cluster also showed increased levels of regulatory T cells (Tregs) and myeloid-derived suppressor cells (MDSCs), suggesting a coexistence of immune activation and suppression within the tumor microenvironment.

Figure 3.

Immune microenvironment and gene expression differences among pancreatic cancer subtypes. (A) Heatmap showing immune cell infiltration scores calculated using multiple immune estimation tools (e.g., MCPcounter and CIBERSORT). Red indicates high infiltration, and blue indicates low infiltration. (B) PCA plot reveals distinct clustering of immune profiles among Cluster A (blue), B (orange), and C (red). (C) Box plots compare specific immune cell infiltration levels, including activated CD8+ T cells, NK cells, and Tregs, across the three subtypes. (D) Volcano plots of differentially expressed genes (DEGs) between subtypes. Red indicates upregulated genes, and blue indicates downregulated genes (P < 0.05, log2FC > 1). (E) GO enrichment analysis reveals significant biological processes (e.g., hormone secretion and insulin secretion), cellular components (e.g., extracellular matrix), and molecular functions (e.g., serine hydrolase activity) enriched in DEGs. (F) KEGG pathway analysis highlights significant enrichment in metabolic pathways (e.g., pancreatic secretion and bile secretion), immune-related pathways (e.g., IL-17 signaling), and drug metabolism (e.g., Cytochrome P450).

Following the immune infiltration analysis that revealed differences in immune characteristics between subtypes, we further explored differential gene expression and functional enrichment among the three metabolic subgroups at the molecular level. Differential analysis of the three subpopulations showed significant gene expression differences among them, and a total of 627 differentially expressed genes (DEGs) were identified, for which we then performed GO and KEGG pathway enrichment analysis (Fig. 3E and 3F). GO analysis showed that DEGs were significantly enriched in biological processes such as extracellular matrix (ECM) organization, cell adhesion, and immune responses. KEGG pathway analysis highlighted significant enrichment in glucose and fatty acid metabolism pathways, underscoring the role of metabolic reprogramming in enabling tumor cells to rapidly acquire energy for proliferation. Additionally, the enrichment of the IL-17 signaling pathway and complement cascade indicates the critical role of inflammatory immunity in the tumor microenvironment, potentially promoting tumor progression through inflammation-driven immune regulation.

Genes associated with pancreatic cancer prognosis identified by univariate Cox regression analysis

Univariate Cox regression analysis was conducted on 627 genes from the TCGA, GSE57495, and combined cohort. In the combined dataset, 234 unfavorable and 196 favorable prognostic genes were identified. Among them, TCGA data set contained 224 unfavorable prognostic and 188 favorable prognostic (Supplemental Digital Content Figure S1D, available at: http://links.lww.com/JS9/H52), while GSE57495 data set contained 48 unfavorable prognostic and 20 favorable prognostics (Supplemental Digital Content Figure S1E, available at: http://links.lww.com/JS9/H52). After their intersection, a total of 47 bad genes and 12 good genes were selected for subsequent modeling (Fig. 4A). ITGB6, CD109, and MET showed significant risk effects in both datasets. In previous reports, we found that ITGB6 promoted ECM degradation by increasing MMP secretion through integrin αvβ6 and activated TGF-β1 to drive cancer progression. CD109 enhanced the migration and proliferation of tumor cells. MET promotes tumor progression by regulating cell migration, invasion, and angiogenesis. In summary, by Cox regression of DEGs in the three subgroups of pancreatic cancer, we obtained a number of genes significantly associated with prognosis. Some of these high-risk genes accelerate cancer progression by synergistically regulating the tumor microenvironment and cellular behavior, highlighting their potential as key therapeutic targets for pancreatic cancer.

Figure 4.

Univariate Cox regression analysis of differentially expressed genes in pancreatic cancer and construction and evaluation of the random forest prognostic model. (A) Hazard ratios (HR) and P-values for differentially expressed genes in the Combined analysis dataset. Genes with HR >1 are associated with poor prognosis, while genes with HR <1 indicate better survival outcomes. (B) The C-index values of 117 machine learning algorithm combinations applied to TCGA datasets. Random forest (RSF) achieved the highest C-index (0.804). (C) Importance ranking of variables in the random forest model. (D) ROC curves showing the predictive performance of the random forest model in the TCGA training set. The model achieved AUCs of 0.976, 0.972, and 0.950 for 1-, 3-, and 5-year survival predictions, respectively. (E) Time-dependent C-index curves demonstrating the model’s stability and high predictive accuracy over time. (F) Kaplan–Meier survival analysis showing significantly poorer survival in the high-risk group compared to the low-risk group (P < 0.001).

Establishment of a prognostic model for pancreatic cancer

To further exploit the key genes obtained, we integrated the use of 10 classical algorithms alone or in combination for modeling, including RSF, LASSO, GBM, Survival-SVM, SuperPC, Ridge Regression, PLSRcox, CoxBoost, Stepwise Cox, and Enet. A total of 117 machine learning combinations were formed. RSF, LASSO, CoxBoost, and stepwise Cox showed reduced and variable screening capabilities. The results show that (Fig. 4B) the RSF model performs best with a C-index of 0.804. The risk genes identified by the RSF model included MET, CA12, and COL17A1, and the protective genes included C10L1, BEX2, and VTN (Fig. 4C).

The predictive value of the model was assessed using ROC curves (Fig. 4D), which demonstrated AUC values of 0.976, 0.972, and 0.950 at 1, 3, and 5 years, respectively, consistent with the time-dependent C-index results (Fig. 4E). Based on the machine learning results, we calculated a RS for each patient and categorized them into high-risk and low-risk groups accordingly. The KM survival curve (Fig. 4F) demonstrated that patients in the high-risk group had significantly shorter survival times compared to those in the low-risk group, further confirming the model’s prognostic accuracy.

Validation and performance evaluation of the random forest model

To validate the predictive performance of the random forest model, the GSE62465 cohort (Fig. 5A) and the combined cohort (Fig. 5B) were used as validation datasets. The median of the RS was used as the boundary value, and the patients were divided into high-risk group and low-risk group for the validation of the model. In the GSE62465 cohort, the accuracy of short-term prediction was higher within 1 or 3 years, with AUC values of 0.784 and 0.708, respectively, whereas the accuracy of long-term prediction (5 years) decreased slightly. Survival analysis showed that the survival rate of the high-risk group was significantly lower than that of the low-risk group (P = 0.00068), and the survival curve was clearly distinguished between the two groups, which further verified the accuracy of the model. In the merged cohort, the model demonstrated excellent performance, with stable and high ROC curves and time-dependent AUC curves, achieving AUC values of 0.975, 0.960, and 0.940 at 1, 3, and 5 years, respectively. Overall, the random forest model of key genes associated with metabolic pathways demonstrated strong risk stratification and survival prediction capabilities, offering a valuable tool for the prognostic assessment of patients with cancer.

Figure 5.

Validation of the random forest prognostic model in external datasets. (A) ROC curves and time-dependent AUC curves in the GSE62465 validation cohort. The AUC values for 1-, 3-, and 5-year survival predictions were 0.784, 0.702, and 0.522, respectively, demonstrating high short-term prediction accuracy despite a decrease in long-term accuracy. Kaplan–Meier survival analysis shows a significantly poorer survival in the high-risk group compared to the low-risk group (P = 0.00068). (B) ROC curves and time-dependent AUC curves in the merged cohort. The model showed high and stable predictive performance, with AUCs of 0.919, 0.889, and 0.859 for 1-, 3-, and 5-year survival predictions, respectively. Kaplan–Meier survival analysis revealed a significantly shorter survival in the high-risk group compared to the low-risk group (P < 0.0001), validating the model’s robustness and clinical utility.

Association of RS with prognosis and clinical characteristics

To deepen the clinical significance of our constructed RS, we analyzed the relationship between the RS of genes and the clinical characteristics of patients, including survival status, tumor grade, and T stage. In the low-risk group, 67% of patients survived, while 33% died. In contrast, only 14% of patients in the high-risk group survived, with 86% deceased. The RS was significantly higher in deceased patients compared to survivors (P < 2.22e−16) (Fig. 6A). For tumor grade, 77% of patients in the low-risk group had low-grade tumors (G1_G2), while 23% had high-grade tumors (G3_G4). In the high-risk group, the proportion of low-grade tumors decreased to 65%, while high-grade tumors increased to 35%. RSs were significantly different between low-grade (G1_G2) and high-grade (G3_G4) tumors (P = 0.0076) (Fig. 6B). Regarding the T stage, 77% of patients in the low-risk group were in the early stage (T1_T2), while 23% were in the advanced stage (T3_T4). Conversely, in the high-risk group, the proportion of advanced-stage patients significantly increased to 89%, while early-stage patients decreased to 11%. RSs were significantly higher in advanced-stage (T3_T4) patients compared to early-stage (T1_T2) patients (P = 0.016) (Fig. 6C). In conclusion, the model’s RS demonstrated strong predictive performance for patient prognosis and robust associations with clinical characteristics.

Figure 6.

Association between the risk score (RS) and clinical characteristics and gene expression. (A) Relationship between RS and survival status. In the low-risk group, 67% of patients survived, while only 14% survived in the high-risk group. (B) Relationship between RS and tumor grade. (C) Relationship between RS and T stage. RSs were significantly higher in advanced-stage patients compared to early-stage patients (P = 0.016). (D) Heatmap of the top 50 positively correlated genes with RSs. (E) Heatmap of the top 50 negatively correlated genes with RSs. (F) Gene Ontology (GO) enrichment analysis of risk-associated genes. (G) KEGG pathway enrichment analysis of risk-associated genes. (H) Reactome pathway enrichment analysis identified significant enrichment in pathways such as mitotic checkpoints, G2/M transition, and DNA replication.

Differential gene expression and functional analysis between high- and low-risk groups

We then performed correlation analysis of all the key genes in the model score and used heatmaps to visualize the top 50 genes positively correlated with prognosis (Fig. 6D) and the top 50 genes negatively correlated with prognosis (Fig. 6E). By GO, KEGG, and Reactome pathway enrichment analysis of high-risk genes (Fig. 6F, 6G, and 6H), the results consistently showed that pathways related to cell cycle (e.g., cell cycle checkpoint, and DNA replication) and tumor proliferation (e.g., p53 signaling pathway and PI3K-Akt pathway) were significantly enriched in the high-risk group. These results indicated that this population of cells had a stronger ability of cell proliferation. In addition, pathways related to metabolic reprogramming, such as lipid metabolism and pyrimidine metabolism, were also highly enriched, further supporting the role of these genes in promoting pancreatic cancer progression.

Development and validation of clinical prediction models

We conducted a univariate Cox regression analysis, which revealed that the RS and risk group (Risk Group) were significantly associated with prognosis (P < 0.001, hazards ratios > 7) (Fig. 7A). Subsequently, a multivariate Cox regression analysis integrating all variables was performed. After adjusting for other factors, the RS (P < 0.001) remained the most significant independent risk factor, confirming its critical role in prognosis prediction (Fig. 7B).

Figure 7.

Risk score (RS) as an independent prognostic factor and nomogram for survival prediction. (A) Univariate Cox analysis shows RS and risk group are significantly associated with prognosis (P < 0.001). (B) Multivariate Cox analysis confirms RS as an independent prognostic factor (P < 0.001). (C) Nomogram integrating RS and clinical variables for predicting 1-, 3-, and 5-year survival, with RS as the strongest predictor. (D) Calibration curves show high consistency between predicted and actual survival, especially for short- and medium-term predictions.

We then constructed a model-and clinical-feature-based nomogram (Fig. 7C) in which the RS remained the most important predictor of survival. While age, stage, grade, and recurrence/metastasis also contributed to survival prediction, their weights were significantly less weighted compared to RS. The calibration curves for the nomogram (Fig. 7D), predicting 1-, 3-, and 5-year survival probabilities, demonstrated excellent alignment with actual survival outcomes for short- and medium-term predictions although some bias were observed in long-term predictions. These results indicated the model’s good calibration performance and robust prognostic prediction, particularly in short- and medium-term survival predictions.

Correlation between RS and immune characteristics

Immunotherapy is currently one of the promising treatment modalities for pancreatic cancer. To explore the relationship between risk genes and immune infiltration, we evaluated the association between immune cell infiltration and RS using multiple estimation methods (including MCP-counter, EPIC, TIMER, etc.) (Fig. 8A). The analysis demonstrated a significant reduction in immune-related metrics, such as the microenvironment score, in the high-risk group, accompanied by a notable decrease in stromal scores. In contrast, epithelial cell levels increased with higher RSs, suggesting a strong negative correlation between the RS and the characteristics of the immune-enhancing microenvironment (Fig. 8B). The tumor microenvironment of high-risk patients was characterized by insufficient immune cell infiltration and reduced stromal cell activity, indicating that these patients were generally in an immunosuppressive microenvironment, further highlighting the key role of RS in predicting the state of the immune microenvironment.

Figure 8.

Association between risk score (RS) and immune cell infiltration and inflammatory factors. (A) Heatmap showing immune cell infiltration levels across high- and low-risk groups, estimated using multiple methods (MCP-counter, EPIC, TIMER, etc.). (B) Correlation analysis of RSs with immune-related scores. RSs are negatively correlated with immune scores and stromal scores, indicating lower immune and stromal activity in the high-risk group. (C) Heatmap of inflammatory factor expression across high- and low-risk groups. (D) Scatter plots showing positive correlations of RSs with PDGFC, IL1B, IL31RA, IL11, EGFR, and IFNE, and negative correlations with CCR7, TGFBR3, IL33, IL17D, IL10RA, CXCL12, CX3CR1, and CCR10.

Inflammatory factors were further divided into four categories: chemokines and their receptors, interleukins and their receptors, interferons and their receptors, and other cytokines. The correlation between RS and inflammatory factors was analyzed. The analysis identified a positive correlation between the RS and gene expression of PDGFC, IL1B, IL31RA, IL11, EGFR, and IFNE and an inverse correlation between the gene expression of CCR7, TGFBR3, IL33, IL17D, IL10RA, CXCL12, CX3CR1, and CCR10 (Fig. 8C). These findings suggest that the tumor microenvironment in high-risk patients is characterized by enhanced inflammation, active angiogenesis, and significant immunosuppression, alongside weakened anti-inflammatory regulation and matrix protection functions. These gene expression changes provide a molecular basis for the RS and highlight potential therapeutic targets for pancreatic cancer.

Targeted drug sensitivity prediction and gene mutation differences

The current study has revealed the associations between the RS, immune microenvironment, and inflammatory factors. To optimize the treatment strategy for high-risk patients, we evaluated the sensitivity of this group of patients to various anticancer agents using oncoPredict. We estimated the IC₅₀ of each cancer agent for each sample and compared the differences in susceptibility to each agent between the high-risk group and the low-risk group (Fig. 9). The IC₅₀ values of AZD6244, AZD8055, and ABT737 were significantly lower in the high-risk group, suggesting that their therapeutic effects were stronger in the high-risk group. These findings provide valuable guidance for the development of precision treatment strategies.

Figure 9.

Drug sensitivity analysis between high- and low-risk groups. Box plots showing the IC₅₀ values of various anticancer drugs in high- and low-risk groups. Drugs such as AZD6244, AZD8055, and ABT737 exhibit significantly lower IC₅₀ values in the high-risk group, indicating higher sensitivity to these drugs in high-risk patients.

The mDNAsi (DNA methylation stemness index) is an indicator reflecting tumor stemness characteristics, derived from the DNA methylation profiles of samples. A higher mDNAsi value indicates a stronger stem cell signature in the sample. Samples were ranked from low to high based on their mDNAsi values (Supplemental Digital Content Figure S2A, available at: http://links.lww.com/JS9/H52). Overall, the mDNAsi values were significantly higher in the high-risk group than in the low-risk group, and the RS was positively correlated with mDNAsi (r = 0.19, P = 0.0176). Similarly, an analysis based on mRNA expression-based stemness index (mRNAsi) showed consistent results (Supplemental Digital Content Figure S2B, available at: http://links.lww.com/JS9/H52). In conclusion, patients in the high-risk group exhibited a stronger tumor stemness signature, which may be associated with increased malignancy and therapy resistance.

We also analyzed the distribution of gene mutations in the high-risk group and the low-risk group. In the high-risk group (73 samples), 66 samples (90.41%) exhibited gene mutations (Supplemental Digital Content Figure S3A, available at: http://links.lww.com/JS9/H52), which was significantly higher than the low-risk group (75 samples), where 43 samples (57.33%) had gene mutations (Supplemental Digital Content Figure S3B, available at: http://links.lww.com/JS9/H52). Notably, the mutation frequencies of KRAS, TP53, CDKN2A, ADAMTS12, and other genes were significantly elevated in the high-risk group (P < 0.05) (Supplemental Digital Content Figure S3C, available at: http://links.lww.com/JS9/H52). These gene mutation characteristics effectively distinguish high-risk from low-risk patients and contribute to accurate predictions of tumor biological behavior.

Discussion

Pancreatic cancer (PAAD) is one of the most malignant tumors, and its incidence and mortality rate are increasing in recent years, so it is urgent to study the molecular mechanism and treatment of pancreatic cancer. Metabolic reprogramming is one of the core processes in the occurrence and development of PAAD. It has been shown that increased glucose uptake and glycolytic flux promote cancer cell proliferation while providing energy for numerous branched biosynthetic pathways^[27,28^]. In PAAD, metabolic reprogramming not only promotes tumor growth but also remodels the immune microenvironment to keep it in an immunosuppressive state. Therefore, a comprehensive understanding of the metabolic changes of pancreatic cancer is of great significance for understanding the mechanism of pancreatic cancer and optimizing the treatment of pancreatic cancer. Compared with previous studies focusing on a single metabolic pathway, this study refined the metabolic pathways of PAAD into three subtypes by using multisource datasets and comprehensively analyzed the relationship between different metabolic pathways and the clinical and molecular characteristics of different metabolic subtypes. We identified 27 metabolic pathways that were significantly associated with survival. Not only did some of these pathways have poor prognosis, we also found that α-linolenic acid and linoleic acid metabolic pathways showed protective effects on the prognosis of pancreatic cancer, which were consistent with previous studies^[29,30^]. Studies have shown that α-linolenic acid can replace arachidonic acid on the cell membrane, thereby reducing angiogenesis, tumor metastasis, and inflammation, and thus exert antitumor effects. This finding highlights that metabolic reprogramming of tumor cells can contribute to tumor progression, providing new therapeutic opportunities for pancreatic cancer by targeting these pathways.

We classified PAAD into three distinct molecular subtypes (Cluster A, B, and C) based on its metabolic profile, which illustrates the metabolic heterogeneity of pancreatic cancer. In order to understand the difference between these three molecular subtypes, we combined the clinical characteristics of the patients for analysis. Cluster C is characterized by enhanced expression of energy and nucleic acid metabolic pathways, and shows aggressive clinical features such as late tumor stage, high recurrence rate, and poor survival. We propose that this isoform reflects a highly proliferative phenotype driven by elevated metabolic activity. In contrast, Cluster A showed protective metabolic features, including upregulation of taurine and hypotaurine metabolic pathways and enrichment of immune response pathways, contributing to a favorable prognosis. Cluster B has an intermediate phenotype that balances invasive and protective metabolic features. In conclusion, the distinction of these three metabolic subtypes provides a robust framework for stratifications of patients with pancreatic cancer for precision therapy.

During pancreatic cancer tumor progression, cancer cells adopt multiple features to evade immune attack, such as disrupting the antigen presentation process or inducing the upregulation of inhibitory immune checkpoint molecules^[31^]. We then analyzed the immune microenvironment of the three subtypes, and as we speculated, there were significant differences in the infiltration of immune cells among the subtypes, with Cluster C showing contradictory characteristics of immune activation and suppression. In Cluster C, the activities of immune killer T cells and NK cells were increased together with the immunosuppressive Tregs and MDSCs, suggesting that the tumor microenvironment of immune escape still exists despite active immune infiltration. Cluster A showed enrichment of immune response pathways, especially the interferon signaling pathway, indicating that this cluster may have stronger antitumor immunity, which is also associated with higher prognosis. These results highlight the interplay between metabolic reprogramming and immune regulation, illustrating the need for combined immunotherapy and metabolic intervention for PAAD.

To quantify prognosis, we adopted a multi-algorithm machine learning strategy and used a combination of multiple algorithms for model construction. The RSF-based risk model achieved excellent prediction performance. The model had high accuracy in TCGA and GSE57495 cohorts (1-, 3-, and 5-year AUC were 0.976, 0.972, 0.950, respectively), which was significantly better than the traditional clinical indicators. By using this model, we identified a series of genes significantly associated with the prognosis of pancreatic cancer. Among them, the role of LAMC2 and MET in pancreatic cancer has been widely reported. LAMC2^[32^] regulates the functional FOSL1-AXL axis through AKT phosphorylation, and LAMC2 inhibition impairs the cell cycle, induces apoptosis, and sensitizes PAAD to MEK1/2 inhibitors (MEK1/2i). MET^[33^] is a growth factor receptor, and its pathway activation is closely related to the invasion, metastasis, and poor prognosis of pancreatic cancer. At present, it has been a research hotspot of targeted therapy drugs. However, there are still some genes whose role in pancreatic cancer has not been studied, and these genes are worthy of further investigation as potential biomarkers and therapeutic targets.

On the treatment side, this study provides a new tool for the precise identification and stratification of patients with pancreatic cancer to guide the selection of populations for immunotherapy and metabolically targeted therapy, and we found increased sensitivity of high-risk tumors to agents such as selumetinib (a MEK inhibitor), AZD8055, and ABT737. It is still challenging to use transcriptome analysis and metabolic analysis in clinical treatment, but the field of cancer diagnosis has developed significantly in recent years, and some cancer diagnostic laboratories have provided transcriptomics-based data to guide treatment decisions^[34^]. With decreasing costs and the standardization of analytical processes, similar metabolic and transcriptomic profiles are expected to be integrated into routine clinical practice in the foreseeable future.

In summary, the present study elucidated the metabolic and molecular complexity of pancreatic cancer by integrating a network of metabolic pathways associated with prognosis and building a model based on differential gene expression. Further validation of these findings should be prioritized in future studies, including modeling of genes significantly associated with pancreatic cancer prognosis and the clinical translation of proposed therapies to improve patient outcomes.

Limitations of this study

Our study has several limitations. First, the lack of experimental validation limits the confirmation of mechanistic roles of identified pathways and genes. The random forest prognostic model we developed has not yet been tested in a prospective clinical setting, which limits its immediate clinical applicability and will be refined in ongoing studies. Future research should aim to address these limitations to improve the translational potential of these findings.

Ethical approval

Not applicable.

Consent

Not applicable.

Sources of funding

This study was supported by several research funding sources, including national and provincial postdoctoral foundations, university-level research funds, and an open project from a provincial clinical immunology laboratory. Specific details have been removed to maintain anonymity during peer review.

Author contributions

H.H.: Data curation and methodology. J.-X.C.: Resources and validation. Wei-Q.C.: Validation. L.-L.Z.: Funding acquisition, writing – original draft. D.-L.Y.: Writing – review and editing.

Conflicts of interest disclosure

The authors declare that they have no competing interests.

Research registration unique identifying number (UIN)

This study does not deal with the relevant issues.

Guarantor

Da-long Yin.

Provenance and peer review

Not commissioned, externally peer-reviewed.

Data availability statement

The data generated and/or analyzed in this study are all publicly available.