Development of molecular markers for authentication of Serra da Estrela cheese

Serra da Estrela (SE) cheese has a Protected Designation of Origin (PDO), and it is manufactured from raw milk of the autochthonous sheep breeds SE and/or Churra Mondegueira (CM). Given the low volume, but high quality, of the milk produced by those breeds, the SE cheese is being subjected to adulteration by the mixture of original milk with cheaper and/or low-quality milk from other more productive breeds, which results in devaluation of the product. Considering the importance of preserving the identity and quality of the SE cheese, this projects goal is to explore and develop methodologies to detect adulterant milk in SE cheese. Molecular methods have been used for cheese authentication given their higher sensitivity and reproducibility than protein-based methods [1]. We previously developed a Random Amplified Polymorphic DNA - Sequence Characterized Amplified Regions (RAPD-SCAR) technique for the differentiation of the adulterant breed Mocha from SE in milk mixtures [2]. Nonetheless, there is still unmet demand for molecular methods able to detect several adulterant breeds, mainly due to low inter-breed genetic variability. The control region of the mitochondrial DNA (mtDNA) has been assigned as a good target for molecular authentication since it is more stable than genomic DNA. In this project, the mtDNA (D-loop) was used to detect cows milk in a mixture with milk from SE, and the differentiation was made based on PCR fragments size. Moreover, bioinformatic analysis of the D-loop region of several breeds revealed population and phylogenetic relationships between Portuguese and foreign sheep breeds, allowing us to increase our understanding of the relatedness between the mtDNA of different sheep breeds. The D-loop has shown potential to provide inter-breed resolution power, since an in-house multiclassification deep learning model, with 16 different classes, was able to properly classify 50 and 100% of SE and CM sequences, correspondingly, in the test dataset. The models predictions for sequences from the remaining breeds in the test dataset have accuracies varying between 0 and 67%. Overall, the model classifies the dataset with 35% accuracy, but with a ROC-AUC of 75%, a commonly used metric that considers the true-positive and falsenegative rates. The in-silico results obtained in this project constitute an important foundation for the development of an integrated experimental approach based on D-loop for SE cheese authentication.This work was supported by ProDOP Serra da Estrela (PDR2020-101-032096) funded by Programa de Desenvolvimento Rural 2014–2020, European Agricultural Fund For Rural Development, and by Portuguese Foundation for Science and Technology under the scope of the strategic funding of UIDB/04469/2020