Multimodal GNNs Enhance Molecular Property Prediction Accuracy

Researchers have developed a novel neural network architecture, the Tri-Branch Modular Fusion Neural Network, designed to improve the accuracy of molecular property prediction. This model integrates three distinct data modalities: the 3D structural information of molecules, their discrete topological grammar represented by SMILES strings, and explicit macroscopic physicochemical descriptors. By combining these diverse data streams, the framework creates a robust multimodal latent space that overcomes limitations of traditional graph neural networks, such as issues with long-range dependencies and oversmoothing. The new architecture was evaluated on the QM9 benchmark, specifically targeting the prediction of atomization energy. It demonstrated superior performance, achieving a Mean Absolute Error of 0.0207 eV with less than one million parameters. This represents a significant 20.6% error reduction compared to strictly geometric baseline models, surpassing sub-chemical accuracy thresholds. The findings suggest that integrating orthogonal data streams offers an efficient alternative to brute-force parameter scaling, providing a powerful surrogate model for high-throughput virtual screening in drug discovery and materials science.