LC-MS/MS-based metabolomics coupled with machine learning for screening candidate biomarkers in bacon

To overcome the subjectivity of traditional bacon authentication, this study developed an objective discrimination framework integrating LC-MS/MS-based metabolomics with machine learning. Metabolomic profiling of bacon samples from pork belly (LRW) and pork rump (LRT) identified 100 differential metabolites common to both groups. The Kruskal-Wallis H test (p < 0.05) identified 75 significant features, which were further reduced to 22 key metabolites through random forest-based feature selection and low-variance filtering. Subsequently, a multidimensional evaluation of these 22 key metabolites (VIF < 75, label correlation >85%, and information gain = 1) highlighted four candidate discriminatory metabolites, including 6-hydroxyoctanoylcarnitine characterizing LRT, and dimethylethanolamine, xanthine, and ethyl hydrogen fumarate associated with LRW. Bidirectional validation using RF, KNN, SVM, and FNN models demonstrated robust discriminatory performance. This framework provides a basis for optimizing bacon production and offers a generalizable paradigm for data-driven biomarker discovery in meat products.