SGF-Net: Fusing SMILES, Graph, and Fingerprints for Molecular Property Prediction

Molecular property prediction is crucial in drug discovery and material design, requiring representations that effectively capture structure-property relationships. Existing multimodal methods are limited by single-architecture graph encoders, which fail to capture both global topology and local bond-level interactions, and by uniform fusion strategies that ignore the distinct semantic roles of different modalities. In this work, we propose SGF-Net, a multimodal molecular representation learning network that integrates SMILES sequences, molecular Graphs, and Fingerprints. Specifically, a dual-channel graph encoder integrating Graphormer and NNConv is designed to generate multiscale representations. Furthermore, a hierarchical bidirectional cross-attention mechanism is adopted to align the features of different modalities at two semantic levels. At the substructure level, cross-attention mutually aligns the SMILES tokens and local graph nodes to capture fine-grained chemical semantics, while at the molecule level, global topology and fingerprints integrate holistic properties. An adaptive gating mechanism further balances modality contributions. Validated by comprehensive ablation studies, SGF-Net achieves competitive performance on six of the eight MoleculeNet benchmarks. The UMAP visualization demonstrates that the fused representation forms significantly more discriminative clusters than single modalities, confirming effective multimodal integration.