Machine Learning for the Cleaner Production of Antioxidant Peptides

Antioxidant peptides (AP) are promising functional foods that have the potential to provide multitude health benefits. They are found in a wide variety of sources, but current methods of discovery and extraction dramatically increases the cost of production which hampers the commercial competitiveness of APs. Focusing on the search and development of short AP sequences that can be easily synthesized through synthetic chemical methods may be able to decrease the cost of production and accelerate lead discovery. However, the traditional method of peptide synthesis that relies on solid-phase chemistry adversely impacts the environment. Thus, minimizing trial-and-error will not only shorten AP discovery but can also make the entire process greener and more cost-effective. In this study, the formulation of a machine learning model that can predict the trolox equivalent antioxidant capacity (TEAC) of tripeptides is presented. It was found that the combination of support vector regression with a polynomial kernel and Blosum indices can accurately predict AP TEAC. The optimized regression model was trained, tested, and externally validated on 121 sequences curated from three different publications. The optimized model demonstrates a 7 % average percent error based on external validation.