EVENT DOSSIER
On September 7, 2026, arXiv released the NEAT-POCKET model, a self-regressive 3D molecular generation method for protein binding pockets. Based on atomic-level generation, maintaining atomic置换 invariance, and explicitly modeling hydrogen atoms, this model demonstrated competitive structure generation performance and significantly faster sampling speeds in benchmark tests on CrossDocked and SPINDR datasets. In addition to full molecule generation, NEAT-POCKET naturally supports pocket condition fragment completion tasks, making it suitable for lead optimization and skeleton extension, and is positioned as a fast, flexible, and practical framework for structure-oriented drug design.
arXiv:2609.05097v1 发布 NEAT-POCKET，这是一种针对蛋白质结合口袋的自回归 3D 分子生成模型。该模型在 CrossDocked 和 SPINDR 数据集上的基准测试显示，其结构基于生成性能具有竞争力，且采样速度显著快于现有基线。NEAT-POCKET 通过原子级生成、保持原子置换不变性并显式建模氢原子来工作，除全分子生成外，还天然支持口袋条件片段补全任务，适用于先导优化和骨架延伸。这些成果将其定位为结构导向药物设计中快速、灵活且实用的框架。