MSGNN: Multi-Scale Graph Neural Network for Molecular Property Prediction
Zhixin Zhang · 2024
In the field of quantum chemistry, predicting molecular properties (such as solubility) is an important research topic that plays a significant role in advancing various fields, including drug design and material discovery. However, traditional methods based on density functional theory (DFT) in physics are time-consuming and less accurate when predicting a large number of molecular properties. Recent studies have shown that Graph Neural Networks (GNNs) can effectively process molecular graphs, thus accomplishing this task more efficiently. However, traditional Graph Neural Networks have certain limitations in information aggregation, as they can only aggregate information from nodes within a specific neighborhood size, and the information obtained by nodes from neighboring nodes may be incomplete. To address this issue, we propose a multi-scale Graph Neural Network model (MSGNN). By setting distance thresholds to obtain neighbor node information at different scales, and then integrating information from different scales for updates, we can more accurately predict the properties of molecules. We have validated the effectiveness of the model on the QM9 dataset, and the experimental results show that compared with a series of previous benchmark models, our method has made significant improvements on multiple indicators.