GSST: Multimodal Graph-SMILES Fusion with Soft SMILES Tokens for Molecular Property Prediction

Accurate prediction of molecular properties is essential for drug discovery. While single modal molecular representations have shown promising results, their generalizability remains limited due to data sparsity and the inherent incompleteness of single modal characterizations. To overcome these limitations, we propose GSST, a novel multimodal pretraining framework that integrates graph-based molecular representations with SMILES sequences through learnable soft SMILES tokens. At the core of GSST is the G2S-Former module, which injects topological information from the graph-based representation into the soft SMILES tokens to enable effective cross-modal interaction while preserving modality-specific features. Extensive experiments on the MoleculeNet and MoleculeACE benchmarks demonstrate that GSST consistently outperforms state-of-the-art methods in molecular property prediction and activity cliff assessment. These results underscore the importance of effective multimodal alignment in capturing shared molecular patterns and alleviating the challenges posed by limited labeled data. GSST represents a scalable and high-throughput approach with significant potential to advance drug discovery.