Virtual Small Molecule Screening And Reaction Yield Prediction Based On Deep Learning

Drug development is a complex project.In order to obtain reliable and reasonable drugs,a large number of experiments are needed,which waste a lot of manpower,material resources and financial resources.In the same manner,in organic synthesis,in order to obtain a target compound or to find high-yielding reaction conditions,it is also necessary to continuously try.From the perspective of economy,environmental protection and efficiency,it is critical to build a variety of reliable computational models for data-driven drug design and organic synthesis.Deep learning,as a machine learning representation algorithm,uses a neural network with multiple layers of nonlinear processing units to learn data representation,and backpropagation algorithm is employed to indicate how the machine should change its internal parameters to discover complex structures in the data set.Deep learning has various frameworks and is widely used in speech recognition,visual recognition and natural language processing,and has achieved excellent results.Due to the accumulation of various omics and biological data,deep learning models have emerged in various fields of drug design,and the performance of deep learning models is better than simple machine learning models in some areas.In addition to drug design,deep learning models with their strong learning and data processing capabilities also perform well in organic synthesis related issues such as reverse synthesis routes prediction and reaction product prediction.The first chapter of this thesis introduces the development history of machine learning and common algorithms,the common frameworks and the training process of deep learning models,and focuses on the specific examples of deep learning methods in some fields of drug design and organic synthesis.The second chapter first introduces the biological background related to the ZAK(sterile alpha motif and leucine zipper containing kinase),and points out that it is an important target for the treatment of cardiovascular diseases and certain cancers.However,there are currently no small molecule inhibitors specifically targeting ZAK.Then the crystal structure of ZAK was introduced,and the binding mode of ligand and ZAK in the complex was analyzed.The ZAK cross-active small molecules that collected from other kinases were docked into its crystal structure,indicating that there is a high risk of false positives in the virtual screening of ZAK small molecules using traditional structure-based drug design methods.In order to improve the enrichment rate of virtual screening,we developed deep learning prediction models.The deep learning classification models were combined with molecular docking for the virtual screening of ZAK small molecule inhibitors.In order to obtain selective molecules for ZAK,we also introduced a kinase spectrum prediction model in the selection process.Using the above methods in combination,we successfully found a selective ZAK small molecule inhibitor with a novel skeleton.In the third chapter,we discuss the application of deep learning in the prediction of reaction yields.We collected high-quality Suzuki-Miyaura reaction data from the literature and used quantification software to calculate the properties of the reactants and catalysts.Using the quantitative properties of the reactants and catalysts and reaction time,reaction temperature and catalyst loading as input of the models,we constructed deep neural network regression models to predict the reaction yields of Suzuki-Miyaura reactions.After the hyper-parameter optimization,the optimal model was determined.The optimal model not only performed well on the modeling data,but also predicted the yields for new reactions that had not been seen.In addition to predicting the reaction yields,our model could also determine high-yielding reaction conditions based on the predicted yields,and all results were experimentally confirmed.In summary,this thesis takes deep learning as the basic means and applies it to two specific and meaningful topics.Our deep learning-based approaches not only enable our projects to achieve good results,but also can be applied to other similar systems,fully affirming that deep learning plays an important role in driving drug design and organic synthesis to achieve data-driven development.