Integrating metabolomics and machine learning with in silico analysis to identify early biomarkers and molecular interactions in sepsis-associated acute kidney injury

Abstract

Abstract Sepsis-associated acute kidney injury (SA-AKI) presents a significant diagnostic challenge in intensive care units (ICUs), largely due to the limitations of current biomarkers. This study utilized early metabolic signatures of sepsis—specifically pre-AKI metabolic features in sepsis—to identify characteristic metabolites capable of predicting the occurrence of SA-AKI within 48 h. Using non-targeted metabolomics, serum samples from 50 sepsis patients were analyzed, including 28 patients in the SA-AKI group and 22 in the sepsis-non-AKI group. Machine learning integration of the least absolute shrinkage and selection operator (LASSO) regression and Boruta algorithms identified diagnostic metabolites. Subsequently, molecular docking was employed to explore potential metabolite-protein interactions. Among 634 detected metabolites, five key biomarkers were identified: Sebacic acid, 1-(β-D-Ribofuranosyl)-1,4-dihydronicotinamide (1-RDN), Threonic acid, Methyl acetate, and Acylcarnitine 10:2. Using leave-one-out cross-validation (LOOCV), where one patient was designated as the test set in each iteration repeated 50 times, the support vector machine (SVM) prediction model achieved an AUC value of 0.89 in the validation cohort. Molecular docking predicted stable binding between 1-RDN and phenylalanine hydroxylase (binding energy = −7.9 kcal/mol), suggesting a potential interaction and crosstalk between fatty acid metabolism and phenylalanine pathway dysregulation. This integrated metabolomics and machine learning approach, complemented by in silico molecular docking, successfully delineated early metabolic signatures of SA-AKI, provided a predictive model for early clinical intervention, and generated testable hypotheses regarding the molecular interactions linking metabolic dysregulation to renal injury.