Paper List

期刊: ArXiv Preprint
发布日期: 2026-03-10
BioinformaticsDL

Equivariant Asynchronous Diffusion: An Adaptive Denoising Schedule for Accelerated Molecular Conformation Generation

Shanghai Academy of Artificial Intelligence for Science, SAIS | Artificial Intelligence Innovation and Incubation (AI3) Institute, Fudan University

Junyi An, Chao Qu, Yun-Fei Shi, Zhijian Zhou, Fenglei Cao, Yuan Qi
Figure
Figure

30秒速读

IN SHORT: This paper addresses the core challenge of generating physically plausible 3D molecular structures by bridging the gap between autoregressive methods (which capture hierarchy but lack global context) and synchronous diffusion models (which offer global conditioning but ignore molecular causality).

核心创新

  • Methodology Proposes Equivariant Asynchronous Diffusion (EAD), a novel framework that assigns independent noise levels to different atoms, enabling asynchronous denoising while maintaining SE(3)-equivariance through graph neural networks.
  • Methodology Introduces a constrained independent sampling strategy during training (Algorithm 1) that reduces combinatorial complexity from O(T^M) to O((2C)^M), making asynchronous diffusion tractable.
  • Methodology Develops a dynamic denoising schedule (Algorithm 2) that uses historical velocity states to adaptively prioritize which atoms to denoise, mimicking hierarchical molecular construction without imposing rigid causal chains.

主要结论

  • EAD outperforms the synchronous denoising baseline EDM (using identical architecture and training iterations) across all metrics, achieving an 8% increase in molecular stability and a 3% improvement in validity.
  • The framework demonstrates that traditional full-molecule diffusion models are special cases of EAD, and the method can be integrated into various diffusion architectures without retraining.
  • Experimental validation shows EAD's ability to generate complete, valid molecules while effectively minimizing cumulative errors that plague autoregressive approaches.
研究空白： Current 3D molecular generation methods face a dichotomy: autoregressive models capture hierarchical relationships but suffer from error accumulation and lack global conditioning, while synchronous diffusion models provide molecule-level control but fail to respect the implicit causal structure of molecular scaffolds and functional groups.

摘要: Recent 3D molecular generation methods primarily use asynchronous auto-regressive or synchronous diffusion models. While auto-regressive models build molecules sequentially, they’re limited by a short horizon and a discrepancy between training and inference. Conversely, synchronous diffusion models denoise all atoms at once, offering a molecule-level horizon but failing to capture the causal relationships inherent in hierarchical molecular structures. We introduce Equivariant Asynchronous Diffusion (EAD) to overcome these limitations. EAD is a novel diffusion model that combines the strengths of both approaches: it uses an asynchronous denoising schedule to better capture molecular hierarchy while maintaining a molecule-level horizon. Since these relationships are often complex, we propose a dynamic scheduling mechanism to adaptively determine the denoising timestep. Experimental results show that EAD achieves state-of-the-art performance in 3D molecular generation.