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

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Molecular property prediction is crucial for efficient drug discovery.
• Current methods using linear SMILES or 2D graphs often fail to capture explicit hierarchical structures and chemical grammar.
• This limitation hinders the development of accurate predictive models.

Purpose of the Study:

• To introduce MolGramTreeNet, a novel multimodal framework designed to improve molecular property prediction.
• To leverage grammar tree-guided representations for a more comprehensive understanding of molecular structures.
• To enhance the accuracy and interpretability of AI-driven drug discovery.

Main Methods:

• Developed MolGramTreeNet, a framework parsing SMILES strings into hierarchical trees using context-free grammar.
• Employed a dual-path encoder: a Transformer for tree-based semantics and a Graph Convolutional Network for 2D topology.
• Fused hierarchical and topological features to generate robust molecular representations.

Main Results:

• MolGramTreeNet demonstrated superior performance across ten diverse datasets.
• Achieved significant performance improvements ranging from 2.86% to 21.47% compared to eight baseline models.
• Ablation studies validated the critical contribution of the grammar tree component.

Conclusions:

• The proposed MolGramTreeNet framework effectively addresses limitations of existing methods by incorporating chemical grammar and hierarchical structures.
• The model offers enhanced accuracy and interpretability, marking a significant advancement in AI-driven drug discovery.
• This approach holds promise for accelerating the identification of novel drug candidates.