Deep learning guided design of dynamic proteins

Deep learning (DL) has advanced the design of static protein structures, but the controlled conformational changes that are hallmarks of natural signaling proteins have remained inaccessible to de novo design. Here, we describe a general DL-guided approach for the de novo design of dynamic changes between intradomain geometries of proteins, similar to switch mechanisms prevalent in nature, with atomic-level precision. Our method involves three general stages: (1) identifying alternative structural states through computational conformational sampling, (2) using DL sequence-to-structure models to restrict the designable sequence space explored during multi-state design, and (3) understanding the molecular basis underlying dynamics through simulations and DL predictions. We solve four structures that validate the designed conformations, demonstrate modulation of the conformational landscape by orthosteric ligands and allosteric mutations, and show that physics-based simulations are in agre..., , , # Data for the deep learning-guided design of dynamic proteins

This dataset contains the plasmid backbone sequences (**Plasmids.zip**), molecular dynamics trajectory data (**MD_files.zip**), design scripts (**Scripts.zip**), and computational structure files (**PDBs.zip**) associated with the publication “Deep learning guided design of dynamic proteins”.

Note: The PDB files have been renumbered to be consistent with our experimentally solved structures deposited in the PDB (i.e. indexed starting from 1 including the 4 residue N-terminal thrombin cleavage site scar - if the scar is not modeled explicitly, then the numbering begins from 5). All data deposited in this repository is numbered according to this convention. However, the single-state and multi-state design scripts using Rosetta/ProteinMPNN use a numbering system where the first residue of the PDB file is indexed as position one (regardless of what residue number is assigned in the PDB file itself), as is standard for Rosetta/...,

深度学习（Deep Learning, DL）已在静态蛋白质结构设计领域取得显著进展，但天然信号蛋白标志性的可控构象变化，至今仍无法通过从头设计（de novo design）实现。在此，我们报道一种通用的深度学习引导方法，可实现蛋白质域内几何结构间动态变化的从头设计，精度可达原子级别，且该设计模拟了自然界中广泛存在的开关机制。
本方法包含三个通用步骤：(1) 通过计算构象采样识别备选结构状态；(2) 利用深度学习序列-结构模型，限制多状态设计过程中可探索的可设计序列空间；(3) 通过模拟与深度学习预测，解析动力学背后的分子机制。我们解析了四个结构以验证所设计的构象，证实正构配体与变构突变可调控构象景观，并证明基于物理原理的模拟与……（原文此处截断）
# 深度学习引导动态蛋白质设计相关数据集
本数据集包含与论文《深度学习引导的动态蛋白质设计》相关的质粒骨架序列（**Plasmids.zip**）、分子动力学轨迹数据（**MD_files.zip**）、设计脚本（**Scripts.zip**）以及计算结构文件（**PDBs.zip**）。
注：本数据集的PDB文件已重新编号，以匹配我们提交至蛋白质数据库（Protein Data Bank, PDB）的实验解析结构，编号规则为：从1开始计数，包含4个残基的N端凝血酶切割位点标记；若未对该标记进行显式建模，则编号从5开始。本仓库中所有数据均遵循此编号规则。不过，使用Rosetta/ProteinMPNN的单状态与多状态设计脚本采用的编号系统为：无论PDB文件自身分配的残基编号为何，PDB文件的首个残基均被索引为位置1，这也是Rosetta工具的通用规范。