Pancreatic adenocarcinoma (PAAD) is one of the most lethal cancers globally, and its resistance to immunotherapy is partly explained by the profound immunosuppression within its tumor microenvironment. A key concept in understanding this immunosuppression is the distinction between immune hot and cold tumors.
Immune hot tumors are characterized by high T-cell infiltration, active antigen presentation, and an inflammatory microenvironment that makes them potentially responsive to checkpoint inhibitors. Immune cold tumors, by contrast, are infiltrated by immunosuppressive cells like regulatory T cells and myeloid-derived suppressor cells, with low effector T-cell activity.
Two recently described forms of regulated cell death, disulfidptosis and pyroptosis, have emerged as potential modulators of tumor immune phenotype. Disulfidptosis involves cell death driven by disulfide stress and SLC7A11 overexpression, while pyroptosis is an inflammatory form of cell death that can activate anti-tumor immune responses by releasing damage-associated molecular patterns.
Despite growing interest in these cell death pathways, no study had systematically linked disulfidptosis and pyroptosis gene expression patterns to immune hot/cold status and long-term prognosis in pancreatic cancer. This study aimed to fill that gap by developing a machine learning-based prognostic signature grounded in these pathways.
Tumor immune phenotypes were classified using CIBERSORT deconvolution of bulk RNA sequencing data to estimate the abundance of 22 immune cell types from the TCGA-PAAD, ICGC-PAAD-AU, and ICGC-PAAD-CA datasets. Samples were then clustered using ConsensusClusterPlus to identify stable immune subtypes.
Hot tumors were defined by high effector T-cell, NK cell, and activated myeloid infiltration, while cold tumors showed dominance of resting immune cells, M2 macrophages, and regulatory T cells. Hot tumors had significantly better overall survival than cold tumors, validating the biological and clinical relevance of this classification.
Weighted Gene Co-expression Network Analysis (WGCNA) was applied to identify gene co-expression modules most correlated with immune hot/cold status. Among 10 identified modules, the pink and turquoise modules were enriched in cold tumor genes while the black module was enriched in hot tumor genes.
Differentially expressed genes between hot and cold tumors totaled 4,620 genes: 2,055 upregulated and 2,565 downregulated in hot compared to cold tumors. These genes were further filtered by intersection with known disulfidptosis and pyroptosis gene sets to identify the most biologically relevant candidates for prognostic modeling.
Rather than selecting a single machine learning algorithm by convention, the study systematically evaluated 137 combinations of different survival modeling approaches. Algorithms tested included random survival forests, lasso regression, elastic net, support vector machines, ridge regression, stepwise Cox regression, partial least squares Cox regression (plsRcox), and others across multiple preprocessing pipelines.
Each combination was evaluated by its ability to predict overall survival in independent validation cohorts using the concordance index (C-index) as the primary performance metric. This exhaustive approach identified the optimal algorithm pair for building the DPIRG prognostic signature.
The best performing combination was survival random forest for feature selection followed by plsRcox (partial least squares Cox regression) for final model construction. This combination achieved the highest C-index values consistently across the three external validation datasets, making it the most generalizable approach.
The comprehensive benchmarking methodology is a significant methodological contribution, demonstrating that algorithm selection has a substantial impact on prognostic model performance and that systematic comparison is preferable to the common practice of reporting results from a single arbitrarily chosen algorithm.
The final DPIRG (Disulfidptosis and Pyroptosis-Related Gene) signature was derived from the genes identified by WGCNA module analysis, differential expression between hot and cold tumors, and the survival random forest feature importance rankings. The signature stratified patients into high-risk and low-risk groups with significant differences in overall survival.
Three genes emerged as particularly important contributors to the DPIRG signature. PLEC (Plectin) encodes a cytoskeletal protein involved in cell structural integrity and has been linked to cancer cell migration and metastatic invasion. Its overexpression in high-risk cold tumors suggests a role in the structural remodeling that enables immune exclusion.
TRPV1 (Transient Receptor Potential Vanilloid 1), a thermosensory ion channel, was identified as a hot tumor-enriched gene associated with inflammatory signaling. Its expression pattern suggests potential roles in tumor-immune cell communication and in regulating the tumor inflammatory state.
ITGB4 (Integrin Beta 4) is a cell adhesion molecule known to promote tumor invasion and is linked to epithelial-mesenchymal transition. Its high expression in cold tumor-associated, high-risk patients connects immune exclusion with invasive phenotype, potentially explaining the co-occurrence of immunosuppression and aggressive disease.
To translate the prognostic signature into therapeutic insights, the study performed molecular docking analysis using DOCK 6.10 to identify drugs with binding affinity for the key DPIRG signature proteins. This computational approach screens potential drug-target interactions based on the three-dimensional structures of both the drug and target protein.
Three drug candidates emerged from molecular docking with the DPIRG biomarker proteins. Thalidomide, an immunomodulatory agent originally used for multiple myeloma, showed binding affinity for DPIRG targets with known anti-inflammatory and anti-angiogenic properties. SB-431542, a selective TGF-beta receptor inhibitor, showed relevance given TGF-beta's central role in creating immune-cold pancreatic tumor microenvironments.
Bleomycin A2, a chemotherapy agent that induces DNA strand breaks and has immunogenic cell death properties, emerged as a third candidate. Notably, bleomycin-induced cell death shares features with pyroptosis, connecting the drug's mechanism of action back to the foundational biology of the DPIRG signature.
While these computational predictions require laboratory and clinical validation before therapeutic translation, they demonstrate how a machine learning-derived prognostic signature can be used not only for risk stratification but also as a starting point for rational drug repurposing in pancreatic cancer.
This study establishes a multi-layered framework that connects immune cell infiltration patterns, regulated cell death mechanisms, co-expression gene modules, machine learning survival modeling, and drug discovery into a unified approach for understanding and treating pancreatic cancer.
The DPIRG signature demonstrates that disulfidptosis and pyroptosis genes are not merely bystander markers but active biological contributors to the immune phenotype of PDAC. The mechanistic connection between these cell death pathways and immune hot/cold status provides a biologically coherent explanation for their prognostic value.
The use of three independent external datasets (TCGA-PAAD, ICGC-PAAD-AU, ICGC-PAAD-CA) for signature validation provides stronger evidence of generalizability than single-cohort studies. The consistent performance across geographically diverse patient populations suggests the DPIRG signature captures fundamental biology rather than dataset-specific artifacts.
Future studies should validate the DPIRG signature in prospective cohorts, test whether PLEC, TRPV1, or ITGB4 expression can be assessed by clinical immunohistochemistry, and pursue preclinical validation of the three drug candidates in pancreatic cancer models to determine whether targeting these pathways can convert immune cold tumors to a more therapy-responsive hot phenotype.