MMGNN: Multi-level, multi-color graph neural networks for molecular property prediction

The article introduces the Multi-level, Multi-color Graph Neural Network (MMGNN), a novel framework for molecular property prediction that utilizes hierarchical decomposition of molecular graphs into overlapping atom-type-pair-specific subgraphs. MMGNN-2D and MMGNN-3D leverage chemical and geometric information respectively, achieving high benchmark performance on MoleculeNet with MMGNN-2D attaining a macro-average AUC-ROC of 0.838 and MMGNN-3D achieving 0.956 on BBBP. This approach demonstrates that overlapping interaction-specific graph decomposition can enhance the predictive capabilities of models in molecular property tasks, benefiting practitioners in the field.