| With the continuous development of electronic information technology,artificial intelligence has become the theme of this era.Artificial intelligence has become the core driving force of the fourth industrial revolution,and will,like mechanization,electrification,and informationization,eventually penetrate into every corner of every industry.In the last decade,artificial intelligence technology has once again seen an unprecedented burst of growth and prosperity.The rise of machine learning technology,especially deep learning technology,is an important driving force for this boom in artificial intelligence.Artificial neural network has begun to receive a lot of attention as a representative deep learning,and it has obvious advantages in feature extraction and modeling.Traditional shallow learning is used in QSAR to quantify the different substances,where data of model can not be too large and the model performance is unstable.Therefore,it is necessary to apply artificial neural networks to model establishment verification.This paper mainly introduces using the QSAR method to evaluate the safety risks of drugs and environmental pollutants,and establish a more stable model to predict other un-synthesized or un-test compounds.It make a difference in the promotion of safer molecular drug design synthesis and the production of environmentally friendly compounds,mainly include the following chapters:Chapter 1 Introduction.It mainly introduces the basic principles of QSAR method,research steps and experimental content,and summarizes the development of QSAR in recent years and briefly introduces the research work of this thesis.Chapter 2 Using the orthogonal design and DPS uniform design method to simplify the DNN modeling process.Due to the complexity and iterative nature of deep learning,its performance is generally superior to traditional models.However,when using the deep neural network for the QSAR modeling process,the selection of various parameters(number of neurons,hidden layer,transfer function,data set division,number of iterations,etc.)becomes difficult.Also,in order to prevent over-fitting of the model,an interactivetest set is usually added as a standard in the process of selecting parameters.However,having all the data involved in different sets(training,testing,interactive testing)is a huge workload.Therefore,based on the difficulty of selecting DNN parameters,this paper proposes a new method,which applies orthogonal design to experimental design and applies DPS uniform design in the process of dividing data sets.The combination of the two methods greatly reduces the workload and guarantees the reliability of the final model.Chapter 3 Deep Neural Networks for the prediction of the inhibitory concentrations of chloroquine derivativesThe class of chloroquine(CQ)derivatives with a half-maximal inhibitory concentration value reported in 222 different literatures was selected to establish a DNN model,and a total of 128,000 DNN models were established to determine the optimization parameters in order to select a better model.First,222 compounds were uniformly divided into different groups by k-fold,uniformly distributed in the training set and test set,and the molecules were optimized and the corresponding descriptors were calculated.Then use HM to select 10 descriptors and establish MLR,ANN,DNN.The end result shows that the DNN model shows better performance than MLR and ANN.Chapter 4 QSBR Modeling for Predicting the Biodegradation of Organic Pollutants Using Deep Neural NetworksMLR and DNN models were established for 290 common organic compounds,of which HM screened 9 variables and RF screened 8 variables.The combination of orthogonal design method experiment and DPS uniform design data set partitioning was used to establish a total of 400,000 models to represent all 14,000,000 models,which reduced 35 times of workload.Finally,the combination of different variable methods and models shows that RF-DNN is more accurate and stable.Chapter 5 Main conclusions and prospects. |
PDF Full Text Request