Topology-Aware Hierarchical Graph Diffusion Model for Molecular Graph Generation

Abstract

This work introduces THGD, a Topology-Aware Hierarchical Graph Diffusion Model designed to address the challenges of generating large, structurally complex molecules. THGD employs a coarse-to-fine framework that decouples global topology preservation from local atomic refinement, enabling precise generative control and efficient exploration of broader chemical spaces without relying on restrictive, predefined motif vocabularies. Extensive experiments underscore THGD’s superior performance. It robustly preserves complex structural constraints, achieving up to 2 $\times $ × higher scaffold validity than the previous state-of-the-art model in scaffold-constrained generation task. Furthermore, in molecular generation task, THGD excels in generating large molecules with high distribution fidelity, attaining an FCD score of 80.26 on the challenging GuacaMol dataset, effectively matching the diversity of real-world molecular distributions. These results highlight THGD’s potential to advance molecular design for drug discovery and beyond. Our code is available at https://github.com/hers22/THGD .