为提高具有益生特性格氏乳杆菌GU-G23的生物量，获取其高密度培养因子，该研究首先通过单因素实验和Plackett-Burman实验筛选出该菌株主要的生长营养因子为鱼蛋白胨、胰蛋白胨、柠檬酸三铵。以响应面实验设计组合作为机器模型训练样本，采用随机森林回归（Random Forest Regression，RFR）和径向基神经网络（Radial Basis Function Neural Network，RBF）模型对其培养基配方进行预测。以决定系数（R-squared，R2）、平均绝对误差（Mean Absolute Deviation，MAE）、均方误差（Mean-Square Error，MSE）和平均绝对百分误差（Mean Absolute Percentage Error，MAPE）作为模型评价指标，比较认为RBF在该研究中具有更好的预测性能。随后选择RBF神经网络和遗传算法（Genetic Algorithm，GA）结合对培养基主要成分进行了优化。最终获得该菌株培养基的优化配方为：鱼蛋白胨29.89 g/L，胰蛋白胨23.33 g/L，柠檬酸三铵4.34 g/L，蔗糖15.00 g/L，低聚果糖15.00 g/L，乙酸钠5.00 g/L磷酸氢二钾0.40 g/L，七水硫酸镁0.58 g/L，一水硫酸锰0.29 g/L，吐温-80 1.00 g/L。在此培养基条件下，所得样品活菌数达到5.21×109 CFU/mL，是未优化前的4.57倍。该研究对微生物高密度培养优化预测提供了新的方法。

To increase the biomass of probiotic Lactobacillus gasseri GU-G23 and identify factors associated with a high-density culture, the main nutritional factors required using a single-factor experiment and Plackett-Burman design were established. The three key factors comprised fish peptone, tryptone, and triammonium citrate. The training sample for the machine learning model was generated according to the response surface experimental design. The random forest regression (RFR) and radial basis function neural network (RBF) models were used to predict their culture medium formulas. The coefficient of determination (R-squared, R2), mean absolute deviation (MAE), mean-square error (MSE), and mean absolute percentage error (MAPE) were adopted as model evaluation metrics. Comparisons revealed that RBF exhibited superior predictive performance. Subsequently, a combination of RBF neural network and genetic algorithm (GA) was selected to optimize the main components of the culture medium. The optimized formula comprised 29.89 g/L fish peptone, 23.33 g/L tryptone, 4.34 g/L triammonium citrate, 15.00 g/L sucrose, 15.00 g/L xylo-oligosaccharides, 5.00 g/L sodium acetate, 0.40 g/L dipotassium hydrogen phosphate, 0.58 g/L magnesium sulfate heptahydrate, 0.29 g/L manganese sulfate monohydrate, and 1.00 g/L Tween-80. Under this medium composition, the number of viable bacterial cells reached 5.21×109 CFU/mL, which was 4.57 times higher than that before optimization. This study provides a new approach for the optimization of microbial high-density culture medium prediction.

国家自然科学基金面上项目（31871802）