Pseudotargeted metabolomics profiles potential damage-associated molecular patterns as machine learning predictors for acute pancreatitis

Abstract

Acute pancreatitis (AP) is a common gastrointestinal disease characterized by pancreatic cell damage and inflammation. Given the early clinical diagnosis and management challenges, exploring novel analytical frameworks from new orientations for interrogating AP is urgent. The release of damage-associated molecular patterns (DAMPs) and their receptor recognition initiate sterile inflammation, serving as key drivers in the development and progression of AP. Thus, this study aimed to delineate the underlying correlations between alterations in the DAMP profile and the AP state. We have developed a new framework combining potential DAMPs profiles obtained from pseudotargeted metabolomics method with machine learning (ML) models for AP prediction. 2-(1-Piperazinyl) pyrimidine chemical labeling was utilized to provide characteristic fragment ions and improve the quantitative sensitivity of targeted metabolites. A total of 49 potential DAMPs were identified and semi-quantified from collected serum samples (n = 84), positive or negative for APs. For modeling obtained datasets with five different ML algorithms, the support vector machine model was chosen as the optimal model to differentiate with high accuracy, achieving an area under the receiver-operating characteristic curve (AUROC) of 0.944. It also showed a strong performance in an external independent validation set (AUROC: 0.907). Moreover, the model was interpreted using the Shapley Additive exPlanations analysis to specify the important features and identify specific free fatty acids as key contributors. Overall, the novel framework enables high accuracy in predicting the presence of AP status. Meanwhile, it underlines the utility of DAMPs in inflammatory diseases and provides reference values for diagnosing in first-line clinics.