Research On Prediction Of Chemical Molecules Bond Energy Based On Graph Convolutional Neural Network

In recent years,deep learning drives the vigorous development of artificial intelligence,and has made a lot of achievements in data mining,machine translation,natural language processing,recommendation algorithms and other related fields.With the development of chemical informatics,deep learning has also been applied in the field of chemistry.Chemical bond energy is a basic parameter to describe the properties of chemical bond.The stability of compounds and whether chemical reactions can occur can be predicted by bond energy.The accurate calculation of chemical bond energy is one of the important research directions in the field of chemistry.The traditional calculation method adopts density functional theory,which is complex and time-consuming(the calculation time is minute or even hour).Using deep neural network and model training through a large amount of data can not only accurately predict the chemical bond energy and obtain the accuracy equivalent to the calculation of density functional theory,but also shorten the calculation time to milliseconds.In this paper we research the change of energy produced by the fracture and formation of chemical bonds before and after chemical reaction,which is referred to as the change of reaction enthalpy.Chemical molecular properties depend on the spatial configuration of molecules.The molecular spatial structure can be abstracted as a graph.The nodes and edges of the graph describe atoms and chemical bonds respectively.Graph convolution neural network(GCN)is an efficient model for processing graph data,which can effectively extract molecular features.In this paper,GCN model is used to predict chemical bond energy from the property of reaction enthalpy change.The main research contents of this thesis are as follows:(1)Establish multi-layer GCN network model to predict the enthalpy change of reaction.The atomic features and molecular bond features of compound molecules are extracted by rdkit and other tools as node features and edge features respectively,which are input into GCN network.The molecular feature of the compound are added as the global feature to realize the prediction of reaction enthalpy change.(2)Propose multi-layer GCN network model based on gating mechanism.A multi-layer GCN network model based on gating mechanism is proposed.Based on the GCN network model combined with global features,the gating mechanism is introduced into the residual connection between GCN network layers to obtain more accurate feature information.(3)Use data enhancement and migration learning to improve the accuracy of the model.The method of data enhancement is used to expand the data,and the reversibility of the reaction is used to expand the data set.Then,through migration learning,pre training is carried out on the data set with low accuracy,and then the model is migrated to the data set with high accuracy for fine tuning,and finally we obtained accurate prediction model that the error of the result is less than 1kcal/mol.