DihedralsDiff: A Diffusion Conformation Generation Model That Unifies Local and Global Molecular Structures

Significant advancements have been made in utilizing artificial intelligence to learn to generate molecular conformations, which has greatly facilitated the discovery of drug molecules. In particular, the rapid development of diffusion models has led to major progress in the field of molecular generation. However, existing molecular diffusion generative models often treat atoms within a molecule as independent entities, neglecting the interactions between atoms and the local structure of the molecule. Alternatively, some models focus exclusively on the local structure within the molecule, failing to fully account for the overall molecular architecture. This paper presents a method that considers both the overall molecular structure and the local structure within the molecule. We propose a novel molecular diffusion generative model named DihedralsDiff, which addresses the aforementioned issue by introducing the concept of dihedral subgraphs. Specifically, the model employs the DihedralsEncode encoder to transform the molecular graph into a set of dihedral subgraphs, thereby enabling the consideration of the overall molecular conformation without neglecting the local structure within the molecule. The proposed model demonstrates superior performance and higher efficiency across various downstream tasks involving molecular conformation generation. By introducing the concept of dihedral subgraphs, the molecular graph is represented as a set of dihedral subgraphs, enabling a unified diffusion process that accounts for both the global and local molecular structures. This approach captures and generates molecular conformations with significantly fewer noise addition and denoising steps, greatly enhancing the model's accuracy and efficiency.