Prediction And Uncertainty Analysis Of Molecular Properties Based On Graph Convolutional Neural Network

In recent years,convolutional neural networks have developed rapidly and have achieved good results in image recognition,speech recognition and machine translation.But traditional convolutional neural networks are good at processing regular data,such as images and text.There are graph structure data such as transportation network,social network and biological network in real life.Because of this ubiquity,attention began to be paid to the use of deep learning models on graphical data.However,the graph data has the characteristics of irregularity,diversity,and large scale,which make the construction of graph convolutional neural networks have certain limitations.Graph convolutional neural network,as a generalized neural network structure based on graph structure,has become a research hotspot due to its unique computing ability.The graph convolutional neural network is applied to predict the properties of molecules in the field of biochemistry,and its properties are closely related to the structure.First introduce the concept of graph in graph theory to express the molecular structure.Think of the molecular structure as a graph,where the nodes in the graph represent the atoms in the molecule,and the edges represent the bonds.Then bring it into the model to predict the nature.In this paper,the octane number that has an important influence on gasoline quality is used as the research object,and a graph convolutional neural network under the framework of message passing network is used to predict the octane number.Before the molecules are input into the neural network,the unsupervised algorithm Graph2 vec is used to encode the molecular map to automatically obtain the effective molecular information;at the same time,the Selu function is used to replace Relu as the activation function to further improve the accuracy of the graph convolutional neural network model.The evaluation model indicators select two types of evaluation absolute error(MAE)and mean square error(MSE).Finally,because the neural network is a black box model,its prediction results are poor in interpretability,and the Bayes Grad method is used to visualize the substructures that play a role in the prediction.This paper constructs an end-to-end unsupervised graph convolutional neural network model,which automatically learns to acquire molecular features,eliminates the need for manual feature selection,and reduces errors caused by expert experience;at the same time,it can be extended to the actual octane number measured without experiments.Enrich the sample data.