Abstract:The single-factor analysis combined with the response surface method (RSM) is usually applied for bacterial medium optimization. However, the variables predicted by this method are limited, and it is difficult to reflect the complex metabolic network of strains, limiting its ability to optimize the formulation. This study focuses on the patented strain Lactiplantibacillus plantarum LP1Z, known for its high antagonistic efficacy against Helicobacter pylori. On the basis of single-factor optimization, random forest, deep neural network, and gradient boosting decision tree (LightGBM) were used to construct the optimal prediction model of high-density fermentation medium components. The accuracy of this prediction model and its performance compared to RSM were evaluated. The results indicated that the Pearson correlation coefficient and R2 of the LP1Z-LightGBM prediction model were close to 1, and the RMSE was as low as 0.02, with the best prediction efficiency. The optimal formula of high fermentation medium was obtained by LP1Z-LightGBM prediction, which includes glucose 30.00 g/L, MgSO4 0.30 g/L, MnSO4 0.03 g/L, CH3COONa 7.00 g/L, yeast powder 12.00 g/L, C6H5O7(NH4)3 2.00 g/L, K2HPO4 1.00 g/L, and Tween-80 2.00 mL/L. The density of LP1Z cultured in the optimal fermentation media reached 1.61 (OD600), which was significantly higher than that obtained by RSM (P<0.05). In conclusion, compared with the traditional RSM, LP1Z-LightGBM offers superior predictive capability for the optimization of high-density fermentation medium.
