Dual molecular and clinical machine-learning prognostic modeling in pancreatic ductal adenocarcinoma: a chaperone-mediated autophagy-based framework integrating multi-cohort molecular signatures and a single-center clinical nomogram.

BACKGROUND

Pancreatic ductal adenocarcinoma (PDAC) is characterized by marked molecular, cellular, and clinical heterogeneity. Chaperone-mediated autophagy (CMA) supports adaptation to metabolic and environmental stress, but its cell type-specific distribution and prognostic relevance in PDAC remain unclear.

METHODS

Single-cell RNA sequencing data from GSE212966 were analyzed to characterize CMA-related transcriptional states in PDAC and adjacent non-tumor tissues. Bulk transcriptomic data from TCGA-PAAD were used for differential expression analysis, weighted gene co-expression network analysis, and molecular model development, while ICGC PACA-CA and PACA-AU served as independent validation cohorts. Multiple survival machine-learning approaches were compared to establish a CMA-related prognostic model. Hallmark pathway activity, immune infiltration, and predicted drug sensitivity were evaluated between risk groups. KRT19, the highest-weighted model gene, was selected for in vitro validation. In parallel, an independent single-center cohort of 468 patients was analyzed using eight survival machine-learning methods to identify clinical prognostic factors and construct a nomogram.

RESULTS

CMA-related transcriptional activity varied among cell types, with macrophages showing prominent scores and PDAC-derived macrophages exhibiting higher CMA scores than those from adjacent tissues. Integration of TCGA differential expression analysis and WGCNA identified 105 candidate genes. The StepCox [forward] plus random survival forest model showed favorable overall performance, with C-index values of 0.903, 0.678, and 0.733 in the TCGA, PACA-CA, and PACA-AU cohorts, respectively. High molecular risk was associated with enhanced glycolytic, proliferative, and cell cycle-related signaling, increased M0 macrophages, reduced CD8+ T cells, and differential predicted drug sensitivity. KRT19 overexpression promoted PDAC cell proliferation, colony formation, migration, and invasion. In the single-center cohort, N stage, CA125, vascular tumor thrombus, and total bilirubin ranked highest in weighted prognostic importance. The clinical nomogram achieved AUC values of 0.661 and 0.750 for 1- and 3-year overall survival, respectively.

CONCLUSION

This study identified CMA-related cellular heterogeneity, established a molecular prognostic model that retained prognostic discrimination in two independent validation cohorts, demonstrated the functional relevance of KRT19, and developed an independent clinical prediction tool. These molecular and clinical models provide complementary perspectives on PDAC prognosis and warrant further evaluation in matched prospective cohorts.