Research On Water-related Properties Prediction Based On Molecular Dynamics And Machine Learning

Molecular properties prediction is the fundamental and important research.Accurate molecular properties prediction is widely concerned in various application fields of cheminformatics,such as industrial application,environmental engineering and drug design.Recent years,the great success of data-driven methods and the application of machine learning technology have accelerated the cross-fusion of information in various fields,and has also brought new challenges and opportunities for molecular information acquisition.In this thesis,molecular dynamics and machine learning approaches are used to predict the thermodynamic properties of supercritical water under microwave irradiation,the solubility of compound molecules and the free energy of hydration respectively.Firstly,molecular dynamics simulations are performed to obtain micro-information of molecular and atomic interaction during microwave-assisted heating supercritical water.The results show that the temperature of supercritical water changes exponentially with the intensity of microwave electric field.The microwave energy has been converted into kinetic energy and intermolecular potential energy,and about 40% microwave energy are stored as latter.The diffusion of supercritical water is nonlinear fluctuation.Secondly,the quantitative structure property relationship models based on Ada Boost,XGBoost,Light GBM and Cat Boost machine learning algorithm are built to predict aqueous solubility using the e-dragon molecular descriptors as input variables.The results show that the last three models have better prediction performance than Adaboost,and the Catboost model has the best fitting performance and better stability,which has the advantage of accurately predicting the aqueous solubility of compounds.According to the ranking of feature importance given by the model,it is found that octanol-water partition coefficient descriptor is the main factor affecting the solubility of compounds.Finally,the quantitative structure property relationship models are constructed to predict hydration free energy based on Ada Boost,XGBoost,Light GBM and Cat Boost machine learning algorithm using the RDKit molecular descriptors as input parameters.The results show that the prediction performance of Catboost model is better than the other three models.The root mean square error of Catboost model is smaller and the generalization ability is stronger.RDKit molecular descriptors selected in this experiment can effectively characterize the hydration free energy.To conclude,molecular dynamics and machine learning were used to predict the waterrelated properties,which have the theoretical reference and practical significance for molecular properties prediction under extreme environments of corrosive,explosive,high temperature,high pressure and microwave irradiation.And it also helpful in accurate prediction of molecular properties corresponding to large-scale dataset.Meanwhile,this thesis may provide the new idea to predict the physicochemical and biological properties of molecules by using the cross advantages of integrating information science into molecular dynamics.