Machine learning prediction of dual and dose-response effects of flavone carbon and oxygen glycosides on acrylamide formation.

Abstract

The extensive occurrence of acrylamide in heat processing foods has continuously raised a potential health risk for the public in the recent 20 years. Machine learning emerging as a robust computational tool has been highlighted for predicting the generation and control of processing contaminants. We used the least squares support vector regression (LS-SVR) as a machine learning approach to investigate the effects of flavone carbon and oxygen glycosides on acrylamide formation under a low moisture condition. Acrylamide was prepared through oven heating via a potato-based model with equimolar doses of asparagine and reducing sugars. Both inhibition and promotion effects were observed when the addition levels of flavonoids ranged 1-10,000 μmol/L. The formation of acrylamide could be effectively mitigated (37.6%-55.7%) when each kind of flavone carbon or oxygen glycoside (100 μmol/L) was added. The correlations between acrylamide content and trolox-equivalent antioxidant capacity (TEAC) within inhibitory range (R 2 = 0.85) had an advantage over that within promotion range (R 2 = 0.87) through multiple linear regression. Taking ΔTEAC as a variable, a LS-SVR model was optimized as a predictive tool to estimate acrylamide content (R 2 inhibition = 0.87 and R 2 promotion = 0.91), which is pertinent for predicting the formation and elimination of acrylamide in the presence of exogenous antioxidants including flavonoids.