Equivariant Graph Self-Attention Transformer for Learning Higher-Order Interactions in 3D Molecular Structures
Despite their considerable success in multiple fields, studying 3D molecular structures of varying sizes presents a significant challenge in machine learning, particularly in drug discovery, as existing methods often struggle to accurately capture complex geometric relationships and tend to be less effective at generalizing across diverse molecular environments. To address these limitations, we propose a novel Equivariant Graph Self-Attention Transformer, namely EG-SAT, which effectively leverages both geometric and relational features of molecular data while maintaining equivariance under Euclidean transformations. This approach enables the model to capture molecular geometry through higher-order representations, enhancing its ability to understand intricate spatial relationships and atomic interactions. By effectively modeling the radial and angular distributions of neighboring atoms within a specified cutoff distance using Atom-Centered Symmetry Functions (ACSFs), EG-SAT leads to a more nuanced and comprehensive understanding of molecular interactions. We validate our model on the QM9 and MD17 datasets, demonstrating that EG-SAT achieves state-of-the-art performance in predicting most quantum mechanical properties, thus showcasing its effectiveness and robustness in this domain.