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

Molecular graph-language alignment is a key challenge in developing Large Molecular Graph-Language Models (LMGLMs). However, most existing approaches overlook the intrinsic hierarchical structure of molecules. While a few recent studies have begun incorporating hierarchical molecular representations, they typically rely on simple concatenation strategies that limit the models' ability to model structural dependencies across different molecular levels. Moreover, their cross-modal alignment often relies on an additional contrastive learning stage for aligning the molecular graph and text. To address these limitations, we propose HQMol: Hierarchical Fusion and Query-Guided Alignment for Molecular Graph-Language Modeling. Our framework incorporates a hierarchical fusion graph encoder that enables unified modeling of both local and global molecular semantics. Additionally, we introduce a query-guided graph projector to more effectively bridge the semantic gap between modalities. Experimental results across multiple molecule-language benchmarks demonstrate that the proposed hierarchical fusion architecture and query-guided alignment strategy significantly enhance model performance, demonstrating the effectiveness of HQMol in cross-modal molecular understanding and generation.