MMRFN: Multi-Dimensional Molecular Representation Fusion Network for Molecular Property Prediction

Molecular property prediction plays a crucial role in drug development and materials science. Traditional experimental methods are costly, time-consuming, and risky, while deep learning technologies have shown significant potential in enhancing prediction efficiency and accuracy. Different dimensional molecular representations contain different types of information. For example, one-dimensional sequence representations provide rich contextual information, two-dimensional molecular graphs reveal topological structure, and three-dimensional geometric representations contain geometric structural details. However, existing deep learning methods mostly focus on extracting features from a single dimension and fail to fully exploit the complementary information embedded in different dimensional molecular representations. To address this, we propose a multi-dimensional molecular representation fusion network (MMRFN) for molecular property prediction. This model integrates SMILES sequences, 2D molecular graphs, and 3D geometric structures for comprehensive feature extraction. Using RDKit to generate molecular graphs and GEM to derive 3D conformations, the model employs BiLSTM-Transformer for sequence semantics, MSGNN for spatial structure features, and MPNN for topological information. A fusion module then combines these multi-dimensional features into a unified molecular representation, enhancing prediction performance.