Molecular Property Prediction Algorithm Under Graph Neural Network

As the outbreak and spread of COVID-19 continues around the world,the demand for specific drugs to combat the virus is increasing.Similarly,many rare diseases are facing the embarrassment of having no available drugs.Meanwhile,for the pharmaceutical industry,the increasing cost of drug research and development and the decrease in return on investment are significant threats to industry development.With the development of the field of biochemistry,a large amount of data has been accumulated in research and production.How to make good use of this data has become an increasingly important issue in new drug research and development.Molecular property prediction is one of the most important issues in computer-aided new drug development processes,playing an important role in many downstream applications such as drug screening and drug design.Its main purpose is to predict the physical and chemical properties of molecules by using internal molecular information such as atomic properties,atomic numbers,and chemical bonds.This enables researchers to find compounds that meet expected properties among a large number of candidate compounds,ultimately achieving the goal of accelerating drug screening and drug design.In the past,most molecular property predictions were estimated through density functional theory,but its high computational cost makes it difficult to predict molecular properties within the required time frame,which is not efficient for current new drug research and development requirements.Graph Neural Network(GNN),as a deep learning model for non-Euclidean data,is particularly suitable for modeling and analyzing molecular data.In this model,the nodes in the graph can represent the atoms in the molecule,and the edges in the graph can represent the chemical bonds in the molecule.Thus,molecular property prediction becomes a graph regression problem.However,current work still has some problems.Existing GNN methods have issues such as sample scarcity,inability to include certain molecular features(such as molecular structure)in model design,which leads to insufficient model prediction ability.Also,traditional GNN models are not designed for molecular property prediction tasks,which means that stacking different GNN models does not significantly improve model prediction ability.This work addresses the above shortcomings and proposes a novel molecular property prediction algorithm based on the Graph Neural Network framework.The best innovation is the use of virtual edge technology to simulate the interaction between atoms that are not connected by chemical bonds.This technology is integrated with traditional GNN models to improve the shortcomings of previous models.We refers to this model as Virtual Edge Graph Neural Network(VEGNN).Experimental studies on three public datasets for molecular property prediction have shown that virtual edge technology significantly improves the model’s prediction ability,proving that VEGNN outperforms mainstream models in predicting molecular properties.