MolGramTreeNet: A multimodal molecular property prediction model via grammar tree-constrained molecular representation

Molecular property prediction is pivotal for drug discovery. However, existing methods that rely on linear SMILES or 2D graphs often overlook explicit hierarchical structures and chemical grammar constraints. To address this limitation, we propose MolGramTreeNet, a multimodal framework that leverages grammar tree-guided representations. The framework parses SMILES into hierarchical trees using context-free grammar, explicitly encoding chemical rules. The architecture employs a dual-path encoder: a Transformer captures tree-based hierarchical semantics, while a Graph Convolutional Network extracts 2D topological features. These fused features generate comprehensive molecular representations. Experiments conducted across ten datasets demonstrate that MolGramTreeNet outperforms eight baseline models, achieving performance improvements of 2.86%-21.47%. Ablation studies confirm the grammar tree's contribution. This approach significantly enhances both accuracy and interpretability, thereby advancing AI-driven drug discovery.