Prediction Of Reaction Selectivity Based On Physical Organic Chemical Descriptors

In organic chemistry,accurate and rapid evaluation of the reactivity and selectivity of chemical reactions is crucial and challenging for both improving organic chemistry understanding and computer predictions.Despite the great application potential of machine learning in predicting reaction performance,publicly available chemical reaction selectivity data that is easy to use for machine learning is still quite scarce.Meanwhile,the development of accurate,fast and well-understood chemical reaction selectivity prediction models is still slow due to the complex relationship between chemical structure and reaction selectivity.Therefore,it is of great scientific significance and practical value to: create a selectivity database of chemical reactions,explore the selectivity controlling factors of organic chemical reactions,develop chemical reaction descriptors that reveal the implicit relationship between chemical structure and reaction selectivity,and achieve accurate quantitative prediction of reaction selectivity.Based on this,this thesis focuses on the creation of chemical reaction selectivity database,the development of easy-to-understand reaction descriptors and the accurate prediction of chemical reaction selectivity.Details are as follows:1.In the construction of the calculation database and of the reaction energy barrier and regioselectivity of the radical C-H functionalization of aromatic heterocycles and the exploration of the distribution trend,a computational database containing 6114 reaction energy barriers and 9370 regional competing reaction pair energy barrier difference data was established by DFT calculations,and the main factors affecting the selectivity of radical C-H functionalization of aromatic heterocycles were explored based on the distribution trend of the data.2.In the machine learning modeling and prediction of the reaction regioselectivity of radical C-H functionalization of aromatic heterocycles,based on the aforementioned reaction regioselectivity database,accurate prediction of reaction regioselectivity was achieved by physical organic chemical descriptors and random forest algorithm.The model can accurately judge the regioselectivity of the experimental cases with excellent generalization ability,which provides a useful tool for the regioselectivity prediction of radical C-H functionalization reactions of aromatic heterocycles.3.In the study to improve the user-friendliness of machine learning predictive models of reaction regioselectivity for radical C-H functionalization reaction of aromatic heterocycles,the feasibility of a strategy to first rapidly obtain molecular structures from SMILES by molecular force fields and semi-empirical methods,and then use machine learning pre-trained models to compute physicochemical descriptors of molecules was explored.The model trained with the random forest algorithm and the descriptors computed by this strategy maintains accurate prediction capability while greatly reducing the computational cost of the descriptors,which can effectively help in high-throughput virtual screening of target reactions.4.In order to introduce experimentally determined physical-organic parameters to aid in the prediction of reaction performance in real scenarios,a transfer learning framework considering solvent influence was developed in the modeling of rapid and accurate prediction of the Mayr nucleophilicity parameter of nucleophiles.The framework achieved accurate prediction of target properties on tiny data sets using molecular structures obtained from SMILES of nucleophiles via molecular force fields or semi-empirical methods as input,which aids in the generalization of Mayr nucleophilicity parameter that has a high experimental measurement cost.