基于深度神经网络与Box-Behnken响应面法的灵芝总三萜提取工艺优化  

Optimization of extraction technology of total triterpenes from Ganoderma lucidum by deep neural network and Box-Behnken response surface methodology

作　　者：姚泽远 卜原玲 韩伟 YAO Zeyuan;BU Yuanling;HAN Wei(Shanghai Key Laboratory of New Drug Design,Pharmaceutical Engineering and Process of Chemical Engineering Research Center of the Ministry of Education,School of Pharmacy,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学药学院,制药工程与过程化学教育部工程研究中心,上海市新药设计重点实验室,上海200237

出　　处：《南京工业大学学报（自然科学版）》2023年第3期347-354,共8页Journal of Nanjing Tech University(Natural Science Edition)

基　　金：上海市科技兴农重点攻关项目[沪农科创字(2018)第1-1号]。

摘　　要：本文研究超声辅助提取灵芝总三萜工艺的优化方法。在单因素实验确定实验参数的大致范围和中心点的基础上,通过筛选实验设计Plackett-Burman(PB)发现对得率影响显著的3个因素(提取时间、超声功率、提取温度),并使用Box-Behnken响应面法进行工艺优化。结合所有实验数据使用基于Pytorch框架的Python学习程序建立深度神经网络(DNN),运用蒙特卡洛算法进行预测和寻优,同时还比较响应面法和深度神经网络两种建模方法。结果表明:通过深度神经网络可得到灵芝总三萜提取的最佳工艺条件(超声功率360 W、提取时间12 min、提取温度58℃、液固比17 mL/g、乙醇体积分数88%),此时预测的灵芝总三萜得率为1.878%,实际测得灵芝总三萜得率为1.805%。深度神经网络在不增加实验量的情况下对已有数据进行学习,可以预测得到更优的工艺条件,并且深度神经网络的非线性拟合能力相比于响应面法的多项式拟合有更强的数据处理能力,预测结果更准确。The optimization of ultrasound-assisted extraction of total triterpenes from Ganoderma lucidum was studied in this paper.The approximate range and center point of the experimental parameters were determined by single factor experiment,and extraction time,ultrasonic power,extraction temperature were found to have significant influence on the yield by Plackett-Burman(PB)experiment.Box-Behnken response surface methodology was used to optimize the process.To achieve the most appropriate extraction conditions,Pytorch framework was used to establish a deep neural network(DNN)combined with all previous experimental data,and Monte Carlo method was used to optimize theses data.At the same time,the two modelling methods of response surface and deep neural network were compared.The optimal process parameters were as follows:ultrasound power of 360 W,extraction time of 12 min,extraction temperature of 58℃,solid-liquid ratio of 17 mL/g,and ethanol volume fraction of 87.6%.Under those conditions,the predicted total triterpenes from Ganoderma lucidum was 1.878%,and the experimental result was 1.805%.It was found that the deep neural network could predict better process conditions by learning existing data without increasing the number of experiments,and the nonlinear fitting ability of the deep neural network was stronger than the polynomial fitting ability of the response surface method,which improved the accuracy of the prediction results.

关 键 词：灵芝总三萜 Box-Behnken响应面法 深度神经网络 工艺优化 超声波辅助提取 蒙特卡洛算法 

分 类 号：R284.2[医药卫生—中药学]