Machine learning-driven prognostic model based on sphingolipid-related gene signature in pancreatic cancer: development and validation.

Background: Pancreatic cancer, a highly malignant tumor with poor prognosis, lacks effective early diagnosis and treatment strategies. Sphingolipids have emerged as key players in tumorigenesis, with certain sphingolipid-related genes linked to patient survival. This study aims to identify prognostic glycosphingolipid (GSL)-related genes and construct a predictive model to improve survival prediction and guide personalized treatment. By providing potential biomarkers, our findings may enhance clinical decision-making and offer new insights into pancreatic cancer diagnosis and therapy.

Methods: This study utilized 150 pancreatic cancer samples from The Cancer Genome Atlas-Pancreatic Adenocarcinoma (TCGA-PAAD) and 69 from GSE62452 [Gene Expression Omnibus (GEO)] for training and validation. Cox univariate regression identified sphingolipid-related genes with prognostic value. Over 100 machine learning algorithms, including Cox models, support vector machines (SVM), and random forests (RF), were applied to construct an optimal survival prediction model for pancreatic ductal adenocarcinoma (PDAC). Model accuracy was evaluated using the concordance index (C-index). Enrichment, immune infiltration, mutation spectrum, and cell communication analyses were performed to explore sphingolipid mechanisms in pancreatic cancer.

Results: Using 10 machine learning algorithms, we developed over 100 models to predict sphingolipid-related survival in pancreatic cancer. A robust prognostic model was constructed, incorporating three GSL-related genes (MET, GBA2, DEFB1), represented by the equation: weighted score = 0.469 * MET + (-0.357) * GBA2 + 0.103 * DEFB1. The model demonstrated strong predictive performance, with a C-index of 0.854 for overall survival in 150 pancreatic cancer patients from the TCGA database and 0.652 in 69 patients from the GEO validation set. Pathway enrichment analysis revealed that high-risk patients were significantly enriched in oncogenic and immune-related pathways. Mutation spectrum analysis indicated a higher mutation load in high-risk patients, with mutations concentrated in common oncogenic pathways. Immune infiltration analysis showed that the risk score positively correlated with immune-suppressive genes but negatively correlated with immune-killing cell infiltration. Cell communication analysis highlighted elevated activity in the macrophage migration inhibitory factor (MIF) pathway within high-risk groups, associated with tumor proliferation and immune escape. In conclusion, this study establishes a sphingolipid-based prognostic model with significant potential for predicting pancreatic cancer outcomes.

Conclusions: The sphingolipid-based model accurately predicts pancreatic cancer survival and suggests sphingolipids promote tumor progression by mediating immune-suppressive microenvironments, aiding prognostic prediction and personalized treatment.