Machine learning strategy to boost 1-deoxynojirimycin production through fermentation by Bacillus amyloliquefaciens

The iminosugar 1-deoxynojirimycin (1-DNJ) is a pharmaceutically significant natural alkaloid with expanding therapeutic and nutraceutical applications. Its microbial biosynthesis offers a sustainable and scalable alternative to plant extraction. However, process optimization is constrained by the complex, non-linear interactions between nutritional and physicochemical fermentation parameters. Herein, the study aimed to enhance the production of 1-DNJ by Bacillus amyloliquefaciens using a combination of single-factor experiments and artificial neural network (ANN) models. LC-MS/MS analysis confirmed the strain produced 1-DNJ from soluble starch and peptone. A strategy combining single-factor experiments and an ANN machine learning model was then used to enhance 1-DNJ yield. The culture medium obtained through single-factor optimisation had α-glucosidase inhibitory activity of 82.55% and a 1-DNJ titer of 1064 ± 4.64 mg/L. A computationally efficient ANN, constructed with a weighted resilient backpropagation algorithm (rprop+) and six hidden neurons, was then trained on this limited dataset. The model identified refined fermentation conditions, predicting a further increase to 87.47% inhibition and a titer of 1179 ± 8.44 mg/L. Sensitivity analysis of the validated ANN pinpointed inoculum volume and fermentation temperature as the most critical process variables. This machine learning framework is a viable technical approach for optimising 1-DNJ fermentation. The identified key fermentation parameters lay a solid foundation for subsequent research.