Boltz MCP服务器
通过Docker使用Boltz2进行结构和亲和力预测
用于Boltz2蛋白质建模的MCP(模型上下文协议)服务器,配备13个工具:
- 从序列或YAML配置预测蛋白质3D结构
- 预测蛋白质配体结合亲和力
- 通过异步监控提交长时间运行的作业
- 跟踪和检索作业结果
- 验证蛋白质序列和配体SMILES
- 批量处理多个序列
- 预先计算MSA支持快速配体筛选
Docker 快速入门
方法1:从GitHub中提取预构建映像
# Pull the latest image
docker pull ghcr.io/macromnex/boltz_mcp:latest
# Register with Claude Code
claude mcp add boltz -- docker run -i --rm \
--user `id -u`:`id -g` --gpus all --ipc=host \
-v ~/.boltz:/opt/boltz_cache:ro \
-v `pwd`:`pwd` \
ghcr.io/macromnex/boltz_mcp:latest方法2:在本地构建Docker镜像
git clone https://github.com/MacromNex/boltz_mcp.git
cd boltz_mcp
docker build -t boltz_mcp:latest .
claude mcp add boltz -- docker run -i --rm \
--user `id -u`:`id -g` --gpus all --ipc=host \
-v ~/.boltz:/opt/boltz_cache:ro \
-v `pwd`:`pwd` \
boltz_mcp:latest要求:
- 支持GPU的Docker(
nvidia-docker或带有NVIDIA运行时的Docker) - Boltz2型号检查点位于
~/.boltz(首次本地运行时自动下载) - 已安装克劳德代码
Docker标志详解
| 标志 | 目的 |
|---|---|
-i | MCP stdio的交互模式 |
--rm | 退出后取出容器 |
` --user id -u:id -g ` | 以当前用户身份运行(输出文件归您所有) |
--gpus all | 访问所有GPU |
--ipc=host | PyTorch性能的主机IPC命名空间 |
-v ~/.boltz:/opt/boltz_cache:ro | 装载缓存的模型检查点(只读) |
-v path:path | 装载用于输入/输出访问的工作目录 |
预先计算的MSA用于快速筛查
当针对同一蛋白质筛选多个配体时,挂载一个预先计算的MSA目录以跳过缓慢的MSA服务器查询(每次运行节省约5分钟):
claude mcp add boltz -- docker run -i --rm \
--user `id -u`:`id -g` --gpus all --ipc=host \
-v ~/.boltz:/opt/boltz_cache:ro \
-v /path/to/msa_files:/opt/msa:ro \
-v `pwd`:`pwd` \
boltz_mcp:latest然后通过 msa_path="/opt/msa/protein.a3m" 任何预测工具。MSA .a3m 文件在第一次运行时自动生成 --use_msa_server 可以在输出目录中找到 msa/.
______________________________________________________________________
验证安装
claude mcp list
# You should see 'boltz' in the output可用工具(共13个):
simple_structure_prediction/simple_affinity_prediction--同步预测submit_structure_prediction/submit_affinity_prediction/submit_batch_structure_prediction--异步作业get_job_status/get_job_result/get_job_log/cancel_job/list_jobs--作业管理validate_protein_sequence/validate_ligand_smiles--输入验证list_example_data--列出捆绑示例
______________________________________________________________________
用法示例
快速结构预测
Predict the structure of this protein: MVLSEGEWQLVLHVWAKVEADVAGHGQDILIRLFKSHPETLEKFDRFKHLKTEREQESRKEAAG. Save to /path/to/results/.蛋白质配体亲和力
Predict binding affinity between protein MVLSE... and aspirin (SMILES: CC(=O)Oc1ccccc1C(=O)O). Use simple_affinity_prediction.基于预计算MSA的快速配体筛选
Screen these 10 ligands against KRAS using simple_affinity_prediction with msa_path="/opt/msa/kras.a3m" for each.具有潜力的高质量结构
Submit a structure prediction with use_potentials=True for the sequence in /path/to/protein.fasta. Monitor until complete.______________________________________________________________________
CLI使用情况(无MCP)
# Structure prediction
./env/bin/python scripts/structure_prediction.py --sequence "MVLSE..." --output results/out
# Affinity prediction
./env/bin/python scripts/affinity_prediction.py \
--protein-seq "MVLSE..." --ligand-smiles "CC(=O)O" --output results/out
# With pre-computed MSA (skips MSA server, much faster)
./env/bin/python scripts/affinity_prediction.py \
--protein-seq "MVLSE..." --ligand-smiles "CC(=O)O" \
--msa-path /path/to/protein.a3m --output results/out______________________________________________________________________
故障排除
找不到Docker?
docker --version # Install Docker if missingGPU无法访问?
- 确保安装了NVIDIA Docker运行时
- 与核对
docker run --gpus all ubuntu nvidia-smi
模型检查点缺失?
- 首先运行本地预测以自动下载到
~/.boltz - 或设置
BOLTZ_CACHE=/path/to/cache跑步前
未找到克劳德代码?
npm install -g @anthropic-ai/claude-code______________________________________________________________________
后续步骤
看 detail.md 有关可用工具、YAML配置格式、本地设置和性能优化的全面指南。
许可证
麻省理工学院
