Proteina-Complexa学习笔记

蛋白设计

2026-04-30

1. Proteina-Complexa 介绍

蛋白质相互作用建模是蛋白质设计中的核心内容。而机器学习技术的应用彻底改变了这一领域，使其在药物研发等诸多领域都发挥了重要作用。在这种背景下，基于结构的新型配体设计方法，可以被视为一种条件生成模型，或者通过结构预测工具来实现序列优化（“hallucination”）。Proteina-Complexa将两种方法进行结合，作者在现有的基于流的蛋白质结构生成技术基础上进行改进，利用通过计算方法预测得到的单体蛋白质结构之间的相互作用关系，构建了一个名为“Teddymer”的新大型数据集。该数据集包含了大量人工合成的配体-靶标对，可用于模型的预训练。结合高质量的实验用的多聚体结构，作者构建出性能出色的base model。随后利用这一生成的先验模型进行推理阶段的优化，从而将原本独立的生成技术和hallucination的优势结合。Proteina-Complexa在计算型结合体设计领域树立了新的标杆：与现有的生成式方法相比，它的计算机模拟成功率显著更高。此外，测试时间优化策略也使得其在相同的计算资源限制下，性能远远优于以往的各种方法。另外其还展示了界面氢键的优化方法、基于折叠结构的配体设计技术，以及将这些方法应用于小分子靶标和酶设计领域的应用案例。

2. Proteina-Complexa 安装

这里介绍docker安装

git clone https://github.com/NVIDIA-Digital-Bio/Proteina-Complexa
cd Proteina-Complexa
docker build -t proteina-complexa -f env/docker/Dockerfile .

运行，记得把参数换成自己的，默认权重会下载在protein-foundation-models目录的community_models和ckpts下面，因此我用git的文件夹替换了dockerfile内本身提供的protein-foundation-models，这样每次进行执行的时候能够获取权重

docker run --gpus all --rm -it -v /home/kangsgo/install/Protein_data:/workspace/data -v /home/kangsgo/install/Proteina-Complexa:/workspace/protein-foundation-models proteina-complexa

编辑.env_example文件，进行部分修改

# ==============================================================================
# Complexa Environment Configuration
#
# Setup:
#   complexa init              # Creates .env from this template
#   # Edit .env with your values, then:
#   complexa init <uv|docker>  # Generates env.sh for your runtime
#   source env.sh              # Activates the environment
#
# WARNING: .env contains sensitive credentials.
# Do NOT commit .env to version control!
# ==============================================================================

# ==============================================================================
# USER CONFIGURATION — Edit these to match your setup
# ==============================================================================

# Credentials (SENSITIVE — fill in your values)
GITLAB_TOKEN=TOKEN_HERE
WANDB_API_KEY=YOUR_WANDB_KEY
WANDB_ENTITY=YOUR_WANDB_ENTITY
HF_TOKEN=

# Local paths (host-side) — set these to your machine's paths
LOCAL_CODE_PATH=/workspace/protein-foundation-models
LOCAL_DATA_PATH=/workspace/data/PFM_data
LOCAL_CACHE_DIR=${LOCAL_CODE_PATH}/.cache
LOCAL_CHECKPOINT_PATH=${LOCAL_CODE_PATH}/checkpoints

# Custom docker mounts (comma-separated "host_path:container_path" pairs, leave empty for none)
DOCKER_MOUNTS=

# Logging
LOGURU_LEVEL=INFO

# Cluster access
CLUSTER_USER=USER_NAME_HERE

# ==============================================================================
# DOCKER SETTINGS — Typically unchanged
# ==============================================================================

# Registry
REGISTRY=registry.example.com
REGISTRY_USER='$oauthuser'

# Docker image and container
DOCKER_IMAGE=registry.example.com/org/repo:tag
CONTAINER_NAME=proteina-dev
DOCKERFILE_PATH=env/docker/Dockerfile

# Docker-side paths (container-internal)
DOCKER_REPO_PATH=/workspace/protein-foundation-models
DOCKER_DATA_PATH=/workspace/data/PFM_data
DOCKER_PYTHONPATH=/workspace/protein-foundation-models/src
DOCKER_CHECKPOINT_PATH=/workspace/Proteina-Complexa/checkpoints
DOCKER_CACHE_DIR=/workspace/protein-foundation-models/.cache
DOCKER_HF_HOME=/workspace/protein-foundation-models/community_models/ckpts
DOCKER_HF_HUB_CACHE=${DOCKER_CACHE_DIR}/huggingface/hub

# ==============================================================================
# MODEL CHECKPOINTS — Derived from LOCAL_CODE_PATH, rarely need changes
# ==============================================================================

USE_V2_COMPLEXA_ARCH=False

# Community model checkpoints
COMMUNITY_MODELS_PATH=${LOCAL_CODE_PATH}/community_models
ESM_DIR=${COMMUNITY_MODELS_PATH}/ckpts/ESM2
AF2_DIR=${COMMUNITY_MODELS_PATH}/ckpts/AF2
RF3_DIR=${COMMUNITY_MODELS_PATH}/ckpts/RF3
RF3_CKPT_PATH=${RF3_DIR}/rf3_foundry_01_24_latest_remapped.ckpt

# ==============================================================================
# EXTERNAL TOOLS — Runtime-specific paths to tool executables
# ==============================================================================
# Python code reads the base names (FOLDSEEK_EXEC, SC_EXEC, etc.) via os.getenv().
# The base names default to UV paths. Change them to DOCKER_* if needed.
UV_VENV=${LOCAL_CODE_PATH}/.venv

# UV runtime tools (default for local development with .venv)
UV_FOLDSEEK_EXEC=${UV_VENV}/bin/foldseek
UV_RF3_EXEC_PATH=${UV_VENV}/bin/rf3
UV_SC_EXEC=${LOCAL_CODE_PATH}/env/docker/internal/sc
UV_MMSEQS_EXEC=${UV_VENV}/bin/mmseqs
UV_DSSP_EXEC=${LOCAL_CODE_PATH}/env/docker/internal/dssp
UV_TMOL_PATH=${UV_VENV}/lib/python3.12/site-packages/tmol

# Docker runtime tools (set in Dockerfile; also used for SLURM Pyxis)
DOCKER_FOLDSEEK_EXEC=/workspace/protein-foundation-models/bin/foldseek
DOCKER_RF3_EXEC_PATH=/workspace/.venv/bin/rf3
DOCKER_SC_EXEC=/workspace/protein-foundation-models/bin/sc
DOCKER_MMSEQS_EXEC=/workspace/protein-foundation-models/bin/mmseqs
DOCKER_DSSP_EXEC=/workspace/protein-foundation-models/bin/dssp
DOCKER_TMOL_PATH=/workspace/.venv/lib/python3.12/site-packages/tmol

# Active tool paths — Python reads these via os.getenv()
# Default to UV; change to ${DOCKER_*} to switch local runtime
FOLDSEEK_EXEC=${UV_FOLDSEEK_EXEC}
RF3_EXEC_PATH=${UV_RF3_EXEC_PATH}
SC_EXEC=${UV_SC_EXEC}
MMSEQS_EXEC=${UV_MMSEQS_EXEC}
DSSP_EXEC=${UV_DSSP_EXEC}
TMOL_PATH=${UV_TMOL_PATH}
DATA_PATH=${LOCAL_DATA_PATH}

# Active checkpoint path — YAML configs use ${oc.env:CKPT_PATH}
CKPT_PATH=${LOCAL_CHECKPOINT_PATH}
............

执行后执行如下命令下载权重

complexa download

随后可以初始化环境设置:

complexa init
complexa init docker
source env.sh

在执行目录下面新建一个bin文件夹

mkdir bin
cd bin

将https://github.com/cytokineking/FreeBindCraft/tree/master/functions
里面的dssp与sc下载并放入。

#也可以不做
wget https://mmseqs.com/foldseek/foldseek-linux-gpu.tar.gz
wget https://mmseqs.com/latest/mmseqs-linux-gpu.tar.gz

解压后放入bin目录下。至此安装完成。

验证是否安装成功：

# Validate the config resolves without errors
complexa validate design configs/search_binder_local_pipeline.yaml

3.快速开始

# 3. Design binders for PDL1
complexa design configs/search_binder_local_pipeline.yaml \
    ++run_name=pdl1_test \
    ++generation.task_name=02_PDL1
    
 # 4. Check results
complexa status configs/search_binder_local_pipeline.yaml

其他管道类似如下：

# Ligand binder design
complexa design configs/search_ligand_binder_local_pipeline.yaml \
    ++run_name=ligand_test \
    ++generation.task_name=39_7V11_LIGAND

# AME motif + ligand binder scaffolding
complexa design configs/search_ame_local_pipeline.yaml \
    ++run_name=ame_test \
    ++generation.task_name=M0024_1nzy_v3

# Monomer motif scaffolding (indexed mode). Note motif targets not provided
complexa design configs/search_motif_local_pipeline.yaml \
    ++run_name=motif_test \
    ++generation.task_name=1YCR_AA

如果你和我一样，是24G显存，可以发现没法跑动，发现可以通过修改configs/pipeline/binder/binder_generate.yaml 中的dataloader batch_size改为8可以跑动。

TODO：通过杨子辰老师的指导，主要是beam search和FK streeing费显存，可以进行修改。

#蛋白设计 ¹⁰ #Binder ⁵

boltzgen学习笔记

小康文章阅读笔记

小康文章阅读笔记

Proteina-Complexa学习笔记

1. Proteina-Complexa 介绍

2. Proteina-Complexa 安装

3.快速开始

分享