Metabolomics for origin traceability of lamb: An ensemble learning approach based on random forest recursive feature elimination

The origin traceability of lamb is a longstanding concern for consumers. Despite the widespread application of untargeted metabolomics in meat origin traceability, challenges remain in achieving rapid and accurate identification of biomarkers. This study utilized untargeted metabolomics to analyse five breeds of geographical indication lamb, obtaining profile data comprising a total of 4139 metabolites. Using random forest recursive feature elimination, 29 potential biomarkers were initially identified, which showed significant breed-specific and production environment-related variations. Upon further assessment, a refined panel of 14 metabolic biomarkers demonstrated optimal accuracy and robustness in tracing lamb origin. When combined with the Naive Bayes algorithm, these biomarkers yielded the highest classification accuracy among all evaluated machine learning methods. The random forest recursive feature elimination presents a practical approach for handling high-dimensional metabolomics datasets compared to previous analytical methods. These findings also provide valuable insights into the development of machine learning-based biomarker panels, greatly enhancing the breed-specific traceability of lamb.