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

Researchers introduce MMGNN (Multi-level, Multi-color Graph Neural Networks), a novel neural network architecture that decomposes molecular graphs into interaction-specific subgraphs to improve molecular property prediction. The framework demonstrates competitive performance across multiple benchmarks, with variants optimized for topological and geometric molecular representations.

MMGNN represents a meaningful advancement in computational chemistry by addressing a fundamental limitation in existing molecular neural networks: the tendency to mix chemically distinct signals through a single propagation pathway. Traditional message-passing neural networks struggle to separately process different types of molecular interactions, requiring deeper architectures and more parameters to capture complex relationships. The new framework elegantly solves this through hierarchical decomposition, creating atom-type-pair-specific subgraphs that preserve fine-grained molecular information while isolating interaction classes.

The dual-variant approach demonstrates thoughtful architecture design. MMGNN-2D focuses on topological properties derived from covalent bonds, while MMGNN-3D incorporates geometric information including spatial proximity and angular descriptors. This separation allows each variant to specialize in complementary aspects of molecular structure, with results showing MMGNN-3D excelling on tasks requiring spatial reasoning (BBBP dataset achieving 0.956 AUC-ROC) while MMGNN-2D shows strength in classification tasks (0.838 macro-average AUC-ROC).

For the drug discovery and materials science communities, this represents practical progress in computational efficiency and predictive accuracy. Organizations developing molecules computationally could reduce screening costs by improving prediction reliability. The benchmarking methodology using scaffold splits and multiple independent runs provides confidence in reproducibility. Future applications likely extend to protein interaction prediction and materials property discovery, where similar interaction-specific decomposition could yield comparable benefits. The technical contribution positions MMGNN as a reference architecture for specialized molecular modeling tasks.

• →MMGNN decomposes molecular graphs into atom-type-pair-specific subgraphs, preventing signal mixing from different chemical interactions
• →MMGNN-2D and MMGNN-3D variants show complementary strengths in topological vs. geometric molecular representations
• →Performance metrics include 0.838 macro-average AUC-ROC (classification) and 0.803 RMSE on ESOL regression benchmarks
• →Hierarchical framework maintains atom-level resolution while enabling interaction-specific message passing
• →Leave-one-out analysis demonstrates the method's interpretability regarding atom-type-pair sensitivities