State-Space Architectures for Scalable Diffusion-based 3D Molecule Generation

Generative models, particularly diffusion models, have shown significant promise in accelerating molecular therapeutic and material discovery. However, efficiently generating high-quality, large, and complex molecules remains a challenge. Previous approaches often struggle with handling intricate molecular structures, suffer from slow and memory-intensive diffusion processes, and lack effective mechanisms for capturing long-range dependencies in molecular graphs. In this study, we aim to leverage the long-range dependency modeling capability of the State-Space Models (SSMs) to extend its applicability to 3D molecule generation. Additionally, we introduce a framework leveraging a few-step iterative diffusion process based on a Euclidean State-Space Model for efficient molecule generation. Further, by incorporating input-dependent node selection, the model enhances node context reasoning in molecular graphs, addressing the limitation of previous methods in capturing complex structural relationships.