Lung adenocarcinoma is the most common histological subtype of lung cancer, accounting for over 45 percent of all cases, with increasing incidence driven by industrialization and air pollution in developing countries. While targeted therapies based on driver mutations like EGFR, ALK, and ROS1 have improved outcomes for molecularly defined subsets, immune checkpoint inhibitors achieve an overall response rate of only 20 to 30 percent across all LUAD patients, creating an urgent need for better biomarkers to identify likely responders.
Programmed cell death encompasses 14 distinct regulated modes of cellular self-destruction with central roles in tumor biology. These include apoptosis, pyroptosis, ferroptosis, autophagy, necroptosis, cuproptosis, disulfidptosis, parthanatos, entotic cell death, netotic cell death, lysosome-dependent cell death, alkaliptosis, oxeiptosis, and zinc-dependent cell death. Dysregulation of PCD pathways through genetic alterations such as Fas mutations and Bcl-2 overexpression, or epigenetic modifications, enables tumor cells to bypass normal growth restrictions and evade immune-mediated clearance, driving both tumor progression and treatment resistance.
PCD plays a dual role in cancer: both protecting against malignancy and enabling immune evasion depending on context. When properly functioning, PCD eliminates damaged or abnormal cells. But tumor cells co-opt PCD pathway dysregulation to resist therapeutic killing, suppress anti-tumor immune responses, and establish multidrug resistance. The biomarker potential of PCD-associated gene expression signatures for predicting prognosis and immunotherapy response in LUAD remains unexplored across the full 14-type PCD landscape.
The study integrated bulk RNA-seq data from TCGA-LUAD and three GEO validation cohorts for a total of 948 LUAD tumor samples. TCGA-LUAD provided the training dataset with 485 tumors and 51 normal tissues. Validation was performed in GSE31210 (226 tumors), GSE37745 (106 tumors), and GSE42127 (131 tumors). Cross-cancer validation used ICI-treated melanoma and urothelial carcinoma cohorts. Two scRNA-seq datasets (GSE131907 with 58 samples across disease states; GSE207422 with 6 ICI-treated patients) provided cell-level resolution.
Feature selection began with ssGSEA scoring of 14 PCD patterns followed by univariate Cox regression to identify prognostically significant PCD types. Six of the 14 PCD patterns showed significant prognostic effects (p less than 0.05). Univariate Cox regression then narrowed this to three PCD patterns independently associated with LUAD prognosis: alkaliptosis and disulfidptosis as risk factors, and lysosome-dependent cell death as a protective factor. Differentially expressed genes between tumor and normal tissue that overlapped with these three PCD patterns yielded 37 candidate genes for machine learning.
A collection of 101 algorithm combinations was applied using 10 machine learning methods for robust signature construction. The methods included random survival forest, elastic network, Lasso, Ridge, stepwise Cox, CoxBoost, partial least squares regression for Cox, supervised principal components, generalized boosted regression modeling, and survival support vector machine. Each dataset was split into a 70 percent training set and 30 percent test set, with ten-fold cross-validation within the training set for hyperparameter optimization. The test set remained frozen throughout development, used only for final unbiased evaluation.
Experimental validation included multiplex immunohistochemistry on tumor tissues from 38 LUAD patients receiving neoadjuvant PD-1 therapy. Tissues were collected from 2022 to 2024, with 15 patients achieving major pathological response (less than or equal to 10 percent residual tumor) and 23 patients with non-major pathological response. Multiplex immunofluorescence staining with anti-CD56, anti-CD20, and anti-CD3 antibodies quantified NK cell, B cell, and T cell infiltration respectively, providing direct clinical validation of computationally derived predictions.
The SuperPC algorithm achieved the highest C-index and identified 13 candidate genes, which were refined to seven by multivariate Cox regression. The final PCD signature includes BTK, CA9, CTSW, NAPSA, PLA2G3, SLC7A11, and UNC13D. The risk score formula assigns negative weights to BTK and CTSW and NAPSA (protective factors) and positive weights to CA9, PLA2G3, SLC7A11, and UNC13D (risk factors), reflecting their opposing associations with LUAD prognosis. The median risk score was used to stratify patients into high and low risk groups.
The PCD signature significantly predicted prognosis in both the TCGA training set and all three GEO validation datasets. Kaplan-Meier survival analysis confirmed that the high-risk group had significantly worse overall survival in TCGA-LUAD, GSE31210, GSE37745, and GSE42127 (all p less than 0.05). The high-risk group also showed higher T-stage, N-stage, and TNM-stage, indicating that the PCD signature captures the biological changes associated with more advanced disease at diagnosis.
Protein-protein interaction network and enrichment analysis revealed a PCD-immunity-metabolism regulatory module. STRING-based PPI analysis of the seven signature genes, applying a minimum interaction score of 0.7, identified a tightly coordinated network. GO and KEGG enrichment analyses confirmed that these genes participate in immune processes (T cell activation, inflammatory responses) as well as lipid-related metabolic pathways. The convergence of PCD, immune regulation, and lipid metabolism in a single gene network defines what the authors term the PCD-immune-metabolism core regulatory module.
The low-risk PCD group showed dramatically richer anti-tumor immune cell infiltration across 28 immune cell types. ssGSEA analysis confirmed that activated B cells, T cells, NK cells, macrophages, and dendritic cells were all more abundant in the low-risk group. Multiple computational algorithms including CIBERSORT, MCP-counter, and EPIC corroborated these findings. The ESTIMATE algorithm confirmed that both immune scores and stromal scores were significantly elevated in the low-risk group, while tumor purity was significantly higher in the high-risk group, indicating a more immune-excluded tumor architecture.
TIDE algorithm analysis showed the high-risk group had significantly elevated immune escape potential. The high-risk PCD group showed higher TIDE scores and higher Exclusion scores, indicating a greater likelihood of tumor-mediated immune evasion during ICI treatment. Consistent with this, the TIDE-predicted proportion of ICI responders was significantly higher in the low-risk group. Most immune checkpoint molecules were significantly negatively correlated with PCD risk score, suggesting that PCD pathway dysregulation and immune checkpoint expression are mechanistically linked.
The PCD signature predicted ICI outcomes in external cohorts of melanoma and urothelial carcinoma patients. Low-risk PCD patients consistently experienced significantly better clinical outcomes following ICI treatment across three external cancer-type datasets. In the IMvigor210 urothelial carcinoma cohort, the low-risk group was enriched for the inflamed tumor immune phenotype. This cross-cancer generalizability indicates that the biological mechanism captured by the seven-gene PCD signature, connecting cell death programs to anti-tumor immunity, is a conserved fundamental mechanism rather than a LUAD-specific artifact.
The high-risk PCD group is also predicted to be less sensitive to most common chemotherapy agents. pRRophetic-based IC50 predictions showed that the high-risk group had higher IC50 values for seven of eight common chemotherapy drugs including cisplatin, docetaxel, doxorubicin, etoposide, gemcitabine, vinblastine, and vinorelbine. Only paclitaxel did not show this pattern. While these predictions derive from computational models rather than clinical response data, they suggest that high-risk LUAD patients may face both immunotherapy resistance and chemotherapy resistance, representing a challenging clinical subgroup.
Single-cell analysis of 208,659 cells from GSE131907 mapped the distribution of all seven PCD signature genes across 10 cell subpopulations. BTK was predominantly distributed in myeloid and B cells. CA9, NAPSA, and PLA2G3 were almost exclusively distributed in epithelial cells. CTSW was predominantly distributed in T cells. SLC7A11, a key molecule in both ferroptosis and disulfidptosis, was primarily distributed in epithelial and myeloid cells. UNC13D, involved in vesicle secretion and immune diseases, was widely distributed in T cells, myeloid cells, and epithelial cells.
Advanced LUAD tumors showed the highest PCD scores, validating the signature's ability to differentiate disease severity. When comparing PCD scores across different tissue sources including normal lung, early-stage LUAD, advanced-stage LUAD, metastatic lymph nodes, pleural effusions, and brain metastases, advanced LUAD samples had the highest aggregate PCD signature scores. This gradient from normal to advanced disease confirms that the signature captures biologically meaningful disease progression rather than being an artifact of technical variation.
CellChat analysis showed more intercellular interactions in the PCD-low group with distinct signaling patterns. In PCD-low tumors, the number and strength of intercellular interactions were significantly greater, with T cells, B cells, and myeloid cells more enriched and functioning as active signal receivers. The PCD-high group showed enrichment of the CXCL12-CXCR4 and CXCL2/CXCL3/CXCL8-ACKR1 signaling axes, which facilitate tumor metastasis and recruit immunosuppressive Tregs and MDSCs while activating cancer-associated fibroblasts that create physical stromal barriers against immune cell infiltration.
NAPSA, already used as a diagnostic marker for LUAD in over 85 percent of cases, showed a novel role as an immunotherapy response predictor. NAPSA had the highest correlation with the overall PCD signature among all seven key genes. KM-plot database queries confirmed that high NAPSA expression was associated with significantly better overall survival and progression-free survival. GSEA analysis revealed that immune activation-related pathways were enriched in high-NAPSA patients, while cell proliferation pathways were activated in low-NAPSA patients, explaining the survival advantage.
High NAPSA expression was associated with enhanced antigen presentation and greater dendritic cell infiltration. Analysis of immune activation pathways and cell infiltration profiles showed that antigen presentation was significantly greater in the high-NAPSA expression group, and multiple DC cell subtypes were more highly infiltrated. This suggests that NAPSA may activate anti-tumor immunity through a pathway involving antigen-presenting cells, though the precise mechanism connecting NAPSA protein function to immune activation remains to be determined experimentally.
In 38 clinical LUAD patients receiving neoadjuvant PD-1 therapy, NAPSA-high patients achieved an 81.8 percent major pathological response rate versus 22.2 percent for NAPSA-low patients. This large absolute difference in MPR rate represents a potentially clinically meaningful predictive relationship. Multiplex immunohistochemistry confirmed that the NAPSA-high group had significantly elevated densities of T cells and NK cells, while B cell infiltration did not differ significantly. These findings directly connect NAPSA expression to the immune cell composition changes that underlie superior immunotherapy outcomes.
NAPSA's existing clinical use as a LUAD diagnostic marker creates a strong foundation for translational implementation. Since NAPSA (also known as NapsinA) is already routinely measured in clinical pathology laboratories as part of the standard LUAD diagnostic workup, repurposing NAPSA expression levels to predict PD-1 therapy efficacy would require no new assay development or infrastructure investment. This ready clinical availability distinguishes NAPSA from most novel biomarkers and substantially reduces the barriers to prospective clinical trial validation and subsequent clinical adoption.
The PCD signature accurately predicts prognosis across four independent LUAD datasets and ICI response in three additional cancer-type cohorts. By integrating 14 PCD patterns through 101 machine learning algorithm combinations across multiomics data, the study identified a seven-gene signature that robustly stratifies patients by both survival outcomes and immunotherapy responsiveness. The cross-cancer validation confirms that the core biology captured, linking PCD to tumor microenvironment immune activation, is not LUAD-specific but represents a fundamental mechanism of ICI response across multiple tumor types.
NAPSA's dual identity as a diagnostic marker and immunotherapy predictor represents its key clinical strength. The study's most immediately translatable finding is that an already-established clinical biomarker can be repurposed for treatment selection without additional assay development. Given that the precise mechanism connecting NAPSA function to immune activation remains unknown, future studies should investigate whether NAPSA directly promotes antigen presentation or immune activation through cytokine release, or whether its expression is simply a marker of a particular epithelial differentiation state that is inherently more immunogenic.
The study highlights PCD-immunity-metabolism crosstalk as a framework for developing novel combination therapeutic strategies. The enrichment of lipid metabolic pathways alongside immune regulation within the PCD gene network suggests that interventions targeting ferroptosis, disulfidptosis, or other lipid-associated cell death pathways could synergize with immune checkpoint blockade. High-risk patients who show both immunotherapy resistance and chemotherapy resistance represent an unmet clinical need where PCD pathway targeting could open new therapeutic avenues unavailable within the current standard-of-care framework.