PairReg: A method for enhancing the learning of molecular structure representation in equivariant graph neural networks.

Abstract

The 3D structure of molecules contains a wealth of important information, but traditional 3DCNN-based methods fail to adequately address the transformations of rigid motions (rotation, translation, and mapping). Equivariant graph neural networks (EGNNs) emerge as efficient models to handle molecular 3D structures due to their unique mechanisms for capturing topological properties and equivariance to rigid motions. Historically, the optimization of EGNN models has been achieved by incorporating higher-order features to capture more complex topological properties. However, adding higher-order features incurs high computational costs. To address this issue, we explore the mechanism to mitigate the oversmoothing of equivariant graph neural networks and propose a new method (PairReg) to mitigate oversmoothing by utilizing equivariant information, such as coordinates, to enhance the model's performance. We validate the performance of the model using the QM9 dataset and conduct ablation experiments on the rMD17 dataset. The results show that our approach enhances the model's ability to characterize the 3D structure of molecules and offers new insights for enhancing the performance of EGNNs.