Work

Machine learning-optimized metabolic biomarker panel for precision screening of early-stage pancreatic cancer in new-onset diabetes

Year: 2025

Type: article

Abstract: Introduction: New-onset diabetes (NOD) represents a high-risk population for pancreatic ductal adenocarcinoma (PDAC), yet effective early detection tools for this specific subgroup remain an unmet clinical need. Methods: We conducted a prospective serum metabolomic analysis using UHPLC-MS/MS in 133 NOD patients aged >65 years, including 60 with PDAC (PDAC+NOD) and 73 without (NOD). Multivariate analysis (OPLS-DA) and machine learning approaches were employed to identify and optimize a diagnostic metabolic biomarker panel. Model performance was evaluated using a hold-out validation set following TRIPOD-ML guidelines. Results: We identified 62 differentially expressed serum metabolites (P<0.05, FDR-corrected), primarily implicating branched-chain amino acid metabolism, bile acid biosynthesis, and sphingolipid signaling pathways. Notably, significant reductions in one-carbon metabolism-related metabolites (serine, glycine, homocysteine) were observed in PDAC+NOD patients. Feature selection yielded an optimized 5-metabolite panel comprising glycine, L-serine, L-methionine, L-homocysteine, and L-homocystine. This panel demonstrated high diagnostic accuracy with an AUC of 0.853 (95% CI: 0.786-0.920) and 75.0% accuracy in distinguishing PDAC+NOD from NOD patients. Discussion: Our study establishes a foundational metabolic biomarker strategy for precision screening of early-stage PDAC in NOD populations. The dysregulated one-carbon metabolites provide novel mechanistic insights into PDAC pathogenesis and offer actionable targets for clinical assay development. Future validation in multi-center cohorts is warranted to confirm clinical utility.

Source: Frontiers in Endocrinology

DOI: 10.3389/fendo.2025.1684608