AI-Assisted Pancreatic Metabolomics Reveals Metabolic Signatures Associated with β-Cell Dysfunction in Experimental Type 2 Diabetes

Introduction: Type 2 diabetes mellitus (T2DM) is characterized by insulin resistance and progressive β-cell dysfunction. Pancreatic changes remain incompletely characterized. This study investigated pancreatic metabolic signatures in experimental T2DM and evaluated machine-learning approaches for discriminatory features.

Methods: T2DM was induced in male Wistar rats by a high-fat diet followed by low-dose streptozotocin. Animals were divided into healthy controls (n = 10), diabetic rats (n = 10), and metformin-treated diabetic rats (n = 10). Pancreatic tissues underwent untargeted metabolomic profiling using UHPLC-HRMS. PCA, PLS-DA, and FDR-adjusted statistical testing were performed. Random Forest and XGBoost identified discriminatory features. Robustness was assessed using repeated cross-validation and permutation testing, with feature selection within training datasets to minimize information leakage. Pathway enrichment used KEGG and MetaboAnalyst resources.

Results: A total of 1,186 metabolic features were detected, of which 74 showed significant differences after FDR correction (FDR < 0.05). Multivariate analyses indicated differences among groups. Permutation testing assessed whether PLS-DA discrimination exceeded chance. Diabetic tissue showed alterations in glutamate, branched-chain amino acid, sphingolipid, tricarboxylic acid cycle, and oxidative-stress-related pathways. Ceramide and acylcarnitine species and glutamate-related metabolites increased, whereas citrate, succinate, and selected phosphatidylcholine species decreased. Machine-learning analysis identified ceramide C18:0, leucine, glutamate, and succinate as discriminatory features. Cross-validated performance and permutation-based significance supported candidate signatures, requiring validation in larger independent cohorts.

Conclusions: Pancreatic metabolomics with explainable machine learning identified metabolic signatures associated with the diabetic phenotype. Alterations in amino acid, mitochondrial energy, and sphingolipid metabolism may contribute to pancreatic metabolic dysfunction. Identified metabolites are candidate tissue-level signatures and targets, not established clinical biomarkers. Because pancreatic tissue is not routinely accessible for clinical biomarker testing and the study examined a single time point, longitudinal studies and validation in accessible biospecimens and independent cohorts are required to establish temporal relationships, generalizability, and translational utility.