UPLC-QTOF-MS coupled with machine learning to discriminate between NFC and FC orange juice

The substitution of not from concentrate (NFC) orange juice with from concentrate (FC) orange juice occurs in the market, damaging consumers' interests. An evaluation of the authenticity of NFC orange juice is critical. This study aimed to develop an approach using LC-MS-based metabolomics and machine learning to discriminate between NFC and FC orange juice. Combining principal component analysis and orthogonal projection to latent structures discriminant analysis, 11 differential compounds for NFC and FC orange juices discrimination were identified. Among them, limonin and hydroxymethylfurfural were higher in FC than in NFC samples, whereas the remaining nine compounds showed the opposite trend. During processing, concentration was the key step for the formation of the differential compounds. Therefore, these 11 compounds have great potential for discrimination between NFC and thermal concentrated FC orange juice processed by other sterilization methods. Based on these 11 differential compounds, random forest (RF), support vector machine (SVM), and partial least squares discriminant analysis machine models were used to identify NFC and FC orange juices. The SVM model was the most accurate model to discriminate between NFC and FC orange juices, with 100% accuracy for both the training and validation sets. Subsequently, the SVM model was used for commercial sample identification, and one NFC orange juice was mislabeled. Our results demonstrated that untargeted screening coupled with machine learning could be a powerful tool for the discrimination of NFC and FC juice.