HQMol: Hierarchical Fusion and Query-Guided Alignment for Molecular Graph-Language Modeling

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Natural Language Processing

Background:

• Large Molecular Graph-Language Models (LMGLMs) face challenges in aligning molecular graph and text data.
• Existing methods often ignore the inherent hierarchical structure of molecules, limiting their ability to capture complex dependencies.
• Current approaches may use simple concatenation for hierarchical representations and require separate contrastive learning stages for cross-modal alignment.

Purpose of the Study:

• To address the limitations of existing LMGLMs by proposing a novel framework, HQMol.
• To effectively model both local and global molecular semantics by incorporating hierarchical molecular structures.
• To enhance cross-modal alignment between molecular graphs and text.

Main Methods:

• Developed HQMol, a framework featuring a hierarchical fusion graph encoder for unified local and global molecular semantics.
• Introduced a query-guided graph projector to bridge the semantic gap between molecular graphs and text.
• Evaluated the framework on multiple molecule-language benchmarks.

Main Results:

• HQMol significantly enhances model performance in molecule-language tasks.
• The hierarchical fusion architecture effectively models molecular semantics across different levels.
• The query-guided alignment strategy improves cross-modal understanding and generation.

Conclusions:

• HQMol offers an effective solution for molecular graph-language modeling by leveraging hierarchical structures.
• The proposed methods advance cross-modal understanding and generation in the molecular domain.
• HQMol demonstrates superior performance compared to existing approaches on benchmark datasets.