Enhancing molecular property prediction of transformer models with dual graph representation

Abstract Accurate prediction of molecular properties is central to advancing chemistry, materials science, and drug discovery. Machine learning on molecular graphs depends critically on representations that capture the topology and structure of molecules. Here we propose the dual graph transformer (DGT), a self-attention architecture that jointly models atom and bond graphs to achieve comprehensive molecular encodings. DGT fuses atom and bond features, graph topology and structure, and stereogeometric information within its self-attention module for an effective molecular representation. We benchmark DGT across a range of datasets for molecular property prediction, showing that it considerably outperforms the current state of the art. DGT demonstrates performance contributions from its dual graph representation, relative positional and structural encodings, and stereogeometric information incorporation while also offering interpretability at the molecular structural level. We envision DGT advancing molecular machine learning by improving both the prediction accuracy and interpretability of molecular properties.