Molecular Properties Prediction Based On 3D Graph Neural Network

Molecular property prediction is a widely used computational method that has applications in various fields,including drug design,material development,environmental protection,and food safety.It is an important cornerstone of scientific research.Traditional molecular property prediction methods face a trade-off between calculation accuracy and speed,and are not suitable for large molecular systems or for handling large amounts of molecular data.As a result,their predictions are subject to significant errors and uncertainties.With the development of machine learning and deep learning,graph neural networks based on the similarity between graph and molecular structures have gradually become the mainstream approach for molecular property prediction.These methods can better handle the complexity and high-dimensional features of molecular structures,thereby improving prediction accuracy and efficiency.Against this background,this dissertation proposes two molecular property prediction models based on 3D graph neural networks.To address the complexity of chemical structures in molecules,this dissertation proposes a graph neural network model based on heterogeneous message passing mechanisms to predict the chemical properties of organic small molecules and the force field properties of molecular conformations.This model models molecules as graph structures and saves the topology and stereo structure information of chemical molecules between nodes using different types of edges,giving the model the ability to recognize similar molecular structures.During the interaction process,the model divides the molecular graph representation into two subgraphs and designs a heterogeneous message passing mechanism that acts on each subgraph separately.For the topology structure of the graph,the model uses the Principle Neighborhood Aggregation Convolution method to aggregate the 2D molecular graph,while for the stereo geometry information in the graph,the model designs a 3D message passing mechanism that acts on the complete 3D graph.These two methods are iteratively alternated and information fusion is achieved during the node update process in the graph.Experimental results show that this model can adapt to various tasks on different types of datasets and has significant advantages in accuracy and transferability,demonstrating the effectiveness of the proposed approach.To address the issue of the lack of molecular chemical information representation in the graph neural network based on the heterogeneous message passing mechanism,this dissertation improves the network by proposing a graph neural network model with chemical information embedding for molecular property prediction tasks.This model extracts chemical knowledge information from molecules as a supplement to the molecular graph representation and proposes to use bond energy as a continuous value to distinguish chemical structures with the same discrete information.For data without bond energy labels,this dissertation proposes using a bond energy prediction model based on chemical reaction processes to generate bond energy data as input.For the extraction of chemical perspective information,this dissertation uses a Pathfinder Discovery Network Convolution module to generate the graph adjacency matrix related to edge information,replacing the Principle Neighborhood Aggregation Convolution.In addition,this dissertation improves the model’s loss function to optimize the objective function for specific tasks,reducing the difficulty of model training.Experiments on different types of data demonstrate the importance of heterogeneous information and verify the accuracy,robustness and the small sample learning ability of the graph neural network model with chemical information embedding.