Enhancing Molecular Property Predictions by Learning from Bond Modelling and Interactions

Researchers introduce DeMol, a new dual-graph framework for molecular property prediction that explicitly models both atoms and chemical bonds to achieve superior accuracy. The approach addresses limitations of conventional atom-centric models by incorporating bond-level phenomena like resonance and stereoselectivity, establishing new state-of-the-art results across multiple benchmarks.

• →DeMol uses a dual-graph framework with parallel atom-centric and bond-centric channels to better model molecular structures.
• →The approach addresses key limitations of existing models that treat chemical bonds as simple pairwise interactions.
• →Multi-scale Double-Helix Blocks enable learning of complex atom-atom, atom-bond, and bond-bond interactions.
• →The framework achieved state-of-the-art performance on major benchmarks including PCQM4Mv2, OC20 IS2RE, QM9, and MoleculeNet.
• →Results demonstrate the importance of explicitly modeling bond information for accurate molecular property predictions.