Prediction Of Chemical Molecular Properties Based On Neighborhood Interaction Graph Convolutional Neural Network

In recent years,neural network models based on deep learning have developed rapidly,resulting in many deep learning models such as convolutional neural networks,graph neural networks,short and long term memory networks,generative adversarial networks,and recurrent neural networks.Many achievements have been made in computer vision,recommendation systems,and natural language processing.Traditional deep convolutional neural networks are often used to process regular data,such as image data,which mainly consists of multidimensional pixel matrices.However,in real life,there are a large amount of irregular graph structure data,such as protein molecules,social networks,chemical molecules,and transportation networks.This type of graph structure data does not have translation invariance because the adjacency structure of each node in the graph network is unique.Although traditional algorithms based on machine learning can basically meet the requirements of researchers in molecular property prediction tasks.However,the disadvantage is that it’s not possible to directly input the chemical molecular structure into the machine learning model,it is also necessary to build molecular properties by hand making features from the chemical molecular structure through feature engineering.This process largely relies on the experience of researchers and requires a significant amount of time.Graph Convolutional Neural Network(GCN)and Convolutional Neural Network(CNN)have similar functions and have an excellent feature extractor that can be used for feature extraction from large-scale data.The core idea of graph convolution neural networks is to define message aggregation functions through message passing mechanisms on graph structured data.By iteratively executing message aggregation,the embedded features of nodes are aggregated and updated using edge information to generate node features for specific tasks.Due to the fact that biochemical moleculesare natural graph data with three-dimensional spatial structures,when graph convolution neural networks are used to predict chemical molecular properties,it should be noted that the properties of molecules are closely related to their spatial structure.Firstly,the idea of graph data in graph theory is used to represent the atoms and their topological structures in chemical molecules.Then the molecular structure is regarded as a graph network,where the nodes represent atoms and the edges represent chemical bond.Finally,learn the embedding of the entire molecular diagram structure through graph convolutional neural network,and then predict chemical or physical properties.In this paper,based on the prediction of the properties of chemical molecules in drugs,an improved graph convolution neural network is used to extract the characteristics of chemical molecules,and complete the prediction of the properties of chemical molecules.Specifically,in this paper,graph convolution neural networks are lightweight,and multiple graph convolution layers are spliced through a layer combination mechanism to extract global features of chemical molecules.Then,local feature aggregators are used to extract neighborhood interaction features of nodes.The embeddings obtained by the above two aggregators are spliced together by attention coefficient,forming an end-to-end chemical molecular feature extraction model.And the effectiveness of our method was verified on multiple datasets.