HMS-GNN: hierarchical multi-scales graph neural network for multi-conformer molecular point group prediction

At present, deep learning methods for predicting the 3D conformation of molecules based on their 2D structures do not take into account the symmetry or point group of the molecules. This paper proposes a multi-conformational molecular point group prediction model HMS-GNN(Hierarchical Multi-Scales Graph Neural Network). By first predicting the possible point groups of various conformations of a molecule, it facilitates the full consideration of its symmetry in subsequent 3D conformation prediction. This model only relies on the graph information of the molecule and can predict the possible conformations of the molecule. We constructed the SMPG dataset suitable for multi-conformational molecular point group prediction based on the GEOMQM9 dataset and trained the model. The accuracy of the HMS-GNN model is 0.967, the F1 value is 0.792, and the Hamming loss is 0.033. This method improves the accuracy by 0.0284, increases the recall rate by 0.1166, and improves the F1 value by 0.051 compared to the method of calculating the molecular point group based on the coordinates of the molecule calculated by molecular force field (RDKit-Pymatgen method). The constructed HMS-GNN neural network model achieves good performance in multi-conformational molecular point group prediction and can provide a reference for the prediction of molecular 3D conformations.