🧪 ColabFold Demo: 蛋白质结构预测与可视化

一个完整的 AlphaFold2/ColabFold 离线预测与可视化流程演示，
支持单序列、MSA 输入，集成 PDB 3D 结构展示、PAE 热图分析 和 静态图像导出。

---

🌍 项目概述

本项目是HFUT深度学习课程的Ai4Science的16组代码部分，成员包括：刘思思，柳霄，王学林，何艺超，刘明鑫，其中代码构建部分由何艺超完成。 本项目基于 ColabFold（AlphaFold2 的开源实现）构建了一个轻量级的蛋白质结构预测与可视化工作流，包含：

• ✅ 离线运行脚本：无需 Google Colab，可在本地或服务器上运行。
• ✅ 完整预测流程：从 FASTA 输入 → MSA 搜索 → 结构预测 → 结果输出。
• ✅ 交互式 3D 可视化：使用 nglview 在 Jupyter Notebook 中实时查看预测结构。
• ✅ 自动保存 PNG 图像：生成高质量 PDB 和 PAE 静态图用于报告或论文。

---

🧩 功能概览

模块	功能说明	关键技术
run_single_sequence.sh	单序列无 MSA 模式预测	ColabFold + AlphaFold2-ptm
run_single_sequence_with_msa.sh	使用预生成 MSA 进行预测	MMseqs2 / a3m 文件输入
colabfold_visualization.ipynb	结果可视化核心脚本	nglview, seaborn, matplotlib

---

⚙️ 安装与环境配置

1. 依赖安装（推荐使用 Conda）

# 创建虚拟环境
conda create -n colabfold python=3.9
conda activate colabfold

# 安装关键包
pip install colabfold nglview matplotlib seaborn numpy jupyterlab

2. 下载模型与数据库

# 下载 ColabFold 所需参数（models）
cd /home/u2024170925/Ai4Science-Demo/ColabFold_demo/databases
wget https://storage.googleapis.com/alphafold/alphafold_params_2022-12-06.tar
tar -xvf alphafold_params_2022-12-06.tar

# 下载 UniRef50 数据库（用于 MSA 搜索）
wget https://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref50/uniref50.fasta.gz
gunzip /home/u2024170925/Ai4Science-Demo/ColabFold_demo/databases/uniref50.fasta.gz

---

🚀 快速上手

1️⃣ 准备输入文件

确保以下文件存在：

results_single/
├── protein.fasta         # 单个蛋白序列（FASTA 格式）
└── protein_msa.fasta     # 可选：已生成的 MSA（用于更准确预测）

示例 protein.fasta 内容：

>query1
MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSHGSAQVKGHGKKVADALTNAVA

---

2️⃣ 运行预测（两种模式）

模式一：仅用单序列（快速但精度较低）

bash run_single_sequence.sh

模式二：使用 MSA（推荐，提升预测准确性）

bash run_single_sequence_with_msa.sh

✅ 输出将保存在 results_single/output/ 目录下，包括：

• query1_unrelaxed_rank_001_alphafold2_ptm_model_1_seed_000.pdb
• query1_predicted_aligned_error_v1.json

---

3️⃣ 可视化结果（Jupyter Notebook）

启动 Jupyter 并打开 colabfold_visualization.ipynb：

jupyter notebook
colabfold_visualization.ipynb

示例预测如下：

🧰 项目结构

ColabFold_Ai4Science_Demo/
├── ColabFold/                     # 模型及代码依赖子目录
├── databases/                     # 用于 MSA/模型的数据库文件夹
├── results_single/                # 单序列预测结果输入／输出目录
│   ├── output/                    # 输出文件（PDB、PAE、JSON 等）
│   └── …                          # 脚本运行前的输入（如 FASTA／MSA）
├── .gitignore
├── README.md                      # 项目说明文档
├── colabfold_visualization.ipynb  # Jupyter Notebook 可视化脚本
├── run_single_sequence.sh         # 单序列预测脚本
├── run_single_sequence_with_msa.sh# 使用 MSA 的预测脚本
└── visual.png                     # 蛋白质预测图

---

📜 版权与引用

本项目基于 ColabFold 开源框架构建，遵循其 MIT 许可协议。

若用于学术发表，请引用原始 AlphaFold 和 ColabFold 文献：

• Jumper et al., Nature (2021): "Highly accurate protein structure prediction with AlphaFold"
• Elnaggar et al., BioRxiv (2022): "ColabFold: Fast and accurate protein structure prediction on Colab"

---

❤️ 致谢

感谢 Sokrypton 提供的 ColabFold 工具链，
以及 AlphaFold 团队推动结构生物学进入 AI 新时代。

英文版：

🧪 ColabFold Demo: Protein Structure Prediction & Visualization

A complete AlphaFold2/ColabFold offline workflow demonstration for protein structure prediction and visualization.
Supports both single-sequence and MSA inputs, integrating 3D PDB visualization, PAE heatmap analysis, and static figure export.

---

🌍 Project Overview

This project is part of the Ai4Science course at HFUT (Hefei University of Technology), developed by Team 16: SiSi Liu, Xiao Liu, Xuelin Wang, Yichao He, and Mingxin Liu.
The code implementation was primarily completed by Yichao He.

This project builds a lightweight workflow for protein structure prediction and visualization based on ColabFold (the open-source implementation of AlphaFold2), including:

• ✅ Offline execution scripts — run locally or on servers, no Google Colab required.
• ✅ Full prediction pipeline — from FASTA input → MSA search → structure prediction → result output.
• ✅ Interactive 3D visualization — view predicted structures in real time using nglview in Jupyter Notebook.
• ✅ Automatic PNG export — generate high-quality PDB and PAE plots for reports or papers.

---

🧩 Function Overview

Module	Description	Core Technologies
run_single_sequence.sh	Predict structure from a single sequence without MSA	ColabFold + AlphaFold2-ptm
run_single_sequence_with_msa.sh	Predict structure using pre-generated MSA	MMseqs2 / A3M input
colabfold_visualization.ipynb	Core visualization notebook	nglview, seaborn, matplotlib

---

⚙️ Installation & Environment Setup

1. Install dependencies (recommended: Conda)

# Create virtual environment
conda create -n colabfold python=3.9
conda activate colabfold

# Install essential packages
pip install colabfold nglview matplotlib seaborn numpy jupyterlab

2. Download models and databases

# Download ColabFold model parameters
cd /home/u2024170925/Ai4Science-Demo/ColabFold_demo/databases
wget https://github.com/sokrypton/ColabFold/releases/download/v1.0/params.tar.gz
tar -xzf params.tar.gz

# Download UniRef50 database (for MSA search)
wget https://ftp.uniprot.org/pub/databases/uniprot/current_release/knowledgebase/idmapping/by_organism/UniRef50.fasta.gz
gunzip UniRef50.fasta.gz

---

🚀 Quick Start

1️⃣ Prepare input files

Ensure the following files exist:

results_single/
├── protein.fasta         # Single protein sequence (FASTA format)
└── protein_msa.fasta     # Optional: pre-generated MSA for higher accuracy

Example of protein.fasta:

>query1
MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSHGSAQVKGHGKKVADALTNAVA

---

2️⃣ Run Prediction (Two Modes)

Mode 1: Single-sequence only (fast but less accurate)

bash run_single_sequence.sh

Mode 2: With MSA (recommended for higher accuracy)

bash run_single_sequence_with_msa.sh

✅ The output will be saved in the results_single/output/ directory, including:

• query1_unrelaxed_rank_001_alphafold2_ptm_model_1_seed_000.pdb
• query1_predicted_aligned_error_v1.json

---

3️⃣ Visualize Results (Jupyter Notebook)

Launch Jupyter and open the visualization notebook:

jupyter notebook
colabfold_visualization.ipynb

Example prediction:

---

🧰 Project Structure

ColabFold_Ai4Science_Demo/
├── ColabFold/                     # Core model and code dependencies
├── databases/                     # Databases for MSA/model parameters
├── results_single/                # Input/output directory for single-sequence predictions
│   ├── output/                    # Output files (PDB, PAE, JSON, etc.)
│   └── …                          # Input files before running (FASTA / MSA)
├── .gitignore
├── README.md                      # Project documentation
├── colabfold_visualization.ipynb  # Jupyter Notebook visualization script
├── run_single_sequence.sh         # Script for single-sequence prediction
├── run_single_sequence_with_msa.sh# Script for MSA-based prediction
└── visual.png                     # Example protein visualization

---

📜 License & Citation

This project is built upon the open-source framework ColabFold under the MIT License.

If used in academic publications, please cite the following works:

• Jumper et al., Nature (2021): "Highly accurate protein structure prediction with AlphaFold"
• Elnaggar et al., BioRxiv (2022): "ColabFold: Fast and accurate protein structure prediction on Colab"

---

❤️ Acknowledgments

Special thanks to Sokrypton for providing the ColabFold toolkit, and to the AlphaFold team for advancing structural biology into the AI era.