This guide details the necessary steps to install and configure AlphaFold 3.0.1 using a Conda Python 3.11 environment, and also provides help with installing the newest AlphaFold 3.0.3 & 3.0.4 releases. It includes the installation of Miniconda, environment creation, dependency installation, repository cloning, model configuration, and the preparation of an execution script. It has been designed to rely solely on conda, without the need for installing any additional packages on the operating system, making it possible to perform this installation on any operating system that supports conda. It is assumed that the appropriate graphics drivers for your Linux distribution are correctly installed. For detailed instructions, please refer to the mini-tutorial.
Once the installation is complete, predicted complexes can be further evaluated with AF_pyDock, our practical workflow for combining AlphaFold model confidence with pyDock energy scoring.
Download and install Miniconda (for x86_64 architecture):
# Download the Miniconda installer
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
# Run the installer
bash miniconda.sh
# Source the .bashrc to update your environment
source ~/.bashrcNote: During installation, you can accept the default settings or customize them according to your preferences.
Note: The AlphaFold 3.0.3 installation instructions have been tested on an RTX 4090.
Create and activate a new conda environment named Alphafold3 with Python 3.11:
# Create the environment
conda create -n Alphafold3 python=3.12 # AlphaFold 3.0.3 / AlphaFold 3.0.4
conda create -n Alphafold3 python=3.11 # AlphaFold 3.0.1
# Activate the environment
conda activate Alphafold3
# Prevent Python from using ~/.local for user-installed packages
# This ensures the environment is fully isolated ("closed")
conda env config vars set PYTHONUSERBASE=intentionally-disabled
# Reactivate the environment for the change to take effect
conda deactivate
conda activate Alphafold3Note: During installation and execution of AlphaFold 3, the
AlphaFold3Conda environment needs to be activated.
Why? SettingPYTHONUSERBASE=intentionally-disabledtells pip and Python to ignore~/.localentirely, ensuring all packages come only from the Conda environment. This prevents interference from globally or user-installed Python packages, keeping the environment reproducible and self-contained.
Install the necessary development tools and dependencies within the conda environment, not on the system OS. Note that HMMER is installed using conda and does not require independent installation or compilation, although there may be tools that cannot be installed in this way.
# Install cmake, gcc, and gxx
conda install -c conda-forge cmake gcc gxx
# Install Boost for Python 3.11 or Python 3.12 and numpy
conda install -c conda-forge boost boost-cpp numpy -y # Boost for Python 3.xx
# Install compression libraries
conda install -c conda-forge bzip2 zstd -y
# Install git and zlib
conda install -c conda-forge git zlib -y
# Install HMMER
conda config --add channels bioconda
conda install -c conda-forge hmmer gsl=2.6 -y
# Install and Upgrade pip within the Alphafold3 environment
conda install pip -y
pip install --upgrade pip # Update pip (specific to the AF3 environment)Install the required Python packages via pip, selecting the appropriate package set based on the AlphaFold version, operating system, system architecture, and accelerator backend.
The latest version, AlphaFold 3.0.4, introduces separate installation options for
- Linux CPU
- Linux with an NVIDIA GPU and CUDA 12
- macOS on Apple Silicon using the JAX Metal/MPS backend
For AlphaFold 3.0.4, first install the common package set and then follow the instructions for the appropriate backend.
pip install \
pandas==2.2.3 matplotlib==3.10.0 \
absl-py==2.3.1 chex==0.1.91 dm-haiku==0.0.16 \
filelock==3.16.1 \
jax==0.10.2 jaxlib==0.10.2 jaxtyping==0.3.5 \
jmp==0.0.4 ml-dtypes==0.5.4 numpy==2.4.1 \
opt-einsum==3.4.0 pillow==12.1.0 rdkit==2025.9.4 \
scipy==1.17.0 tabulate==0.9.0 toolz==1.1.0 \
tqdm==4.67.1 typeguard==2.13.3 \
typing-extensions==4.15.0 zstandard==0.25.0 \
aiofiles==25.1.0 annotated-types==0.7.0 einshape==1.0 \
"etils[epath]==1.13.0" flax==0.12.2 fsspec==2026.6.0 \
humanize==4.15.0 immutabledict==4.2.2 \
importlib-resources==6.5.2 msgpack==1.1.2 \
nest-asyncio==1.6.0 optax==0.2.6 \
orbax-checkpoint==0.11.31 packaging==25.0 \
protobuf==6.33.3 psutil==7.2.1 \
pydantic==2.12.5 pydantic-core==2.41.5 \
pyyaml==6.0.3 qwix==0.1.5 rich==14.2.0 \
markdown-it-py==4.0.0 mdurl==0.1.2 pygments==2.19.2 \
simplejson==3.20.2 tensorboardx==2.6.4 \
tensorstore==0.1.80 tokamax==0.0.12 \
treescope==0.1.10 typing-inspection==0.4.2 \
wadler-lindig==0.1.7 zipp==3.23.0 \
biopython==1.83After installing the common packages, select the appropriate backend below.
No additional Python packages are required. The common installation already includes the CPU versions of:
jax==0.10.2
jaxlib==0.10.2
This option does not install any CUDA or NVIDIA packages.
Use this option for supported NVIDIA GPUs, including Blackwell and earlier CUDA-capable NVIDIA architectures.
Install the CUDA 12 backend and the exact NVIDIA library versions recorded in the AlphaFold 3.0.4 lock file:
pip install \
"jax[cuda12]==0.10.2" jaxlib==0.10.2 \
jax-cuda12-pjrt==0.10.2 "jax-cuda12-plugin[with-cuda]==0.10.2" \
nvidia-cublas-cu12==12.9.1.4 nvidia-cuda-cccl-cu12==12.9.27 \
nvidia-cuda-cupti-cu12==12.9.79 nvidia-cuda-nvcc-cu12==12.9.86 \
nvidia-cuda-nvrtc-cu12==12.9.86 nvidia-cuda-runtime-cu12==12.9.79 \
nvidia-cudnn-cu12==9.17.1.4 nvidia-cufft-cu12==11.4.1.4 \
nvidia-cusolver-cu12==11.7.5.82 nvidia-cusparse-cu12==12.5.10.65 \
nvidia-nccl-cu12==2.29.2 nvidia-nvjitlink-cu12==12.9.86 \
nvidia-nvshmem-cu12==3.5.19Do not install these CUDA packages on macOS or on a Linux system intended to use only the CPU backend.
Use this option on ARM-based macOS systems, including Apple Silicon processors such as the M1, M2, M3, M4, and later generations.
Install JAX together with the Apple Metal/MPS backend:
pip install \
jax==0.10.2 \
jaxlib==0.10.2 \
jax-mps==0.10.9Do not install the NVIDIA CUDA packages on macOS.
AlphaFold 3.0.4 installation summary:
Platform Required commands Linux CPU Common package command only Linux with NVIDIA GPU Common package command + CUDA 12 backend command macOS Apple Silicon Common package command + JAX-MPS command CUDA and JAX-MPS are alternative accelerator backends and must not be installed together.
pip install pandas==2.2.3 matplotlib==3.10.0 absl-py==2.3.1 chex==0.1.91 \
dm-haiku==0.0.16 filelock==3.16.1 \
"jax[cuda12]==0.9.1" jax-cuda12-pjrt==0.9.1 jax-cuda12-plugin[with-cuda]==0.9.1 \
jaxlib==0.9.1 jaxtyping==0.3.5 jmp==0.0.4 ml-dtypes==0.5.4 \
numpy==2.4.1 nvidia-cublas-cu12==12.9.1.4 \
nvidia-cuda-cupti-cu12==12.9.79 nvidia-cuda-nvcc-cu12==12.9.86 \
nvidia-cuda-runtime-cu12==12.9.79 nvidia-cudnn-cu12==9.17.1.4 \
nvidia-cufft-cu12==11.4.1.4 nvidia-cusolver-cu12==11.7.5.82 \
nvidia-cusparse-cu12==12.5.10.65 nvidia-nccl-cu12==2.29.2 \
nvidia-nvjitlink-cu12==12.9.86 opt-einsum==3.4.0 pillow==12.1.0 \
rdkit==2025.9.4 scipy==1.17.0 tabulate==0.9.0 toolz==1.1.0 \
tqdm==4.67.1 typeguard==2.13.3 \
typing-extensions==4.15.0 zstandard==0.25.0 \
aiofiles==25.1.0 annotated-types==0.7.0 einshape==1.0 etils==1.13.0 \
flax==0.12.2 fsspec==2026.1.0 humanize==4.15.0 \
immutabledict==4.2.2 importlib-resources==6.5.2 msgpack==1.1.2 \
nest-asyncio==1.6.0 nvidia-cuda-cccl-cu12==12.9.27 \
nvidia-cuda-nvrtc-cu12==12.9.86 nvidia-nvshmem-cu12==3.5.19 \
optax==0.2.6 orbax-checkpoint==0.11.31 packaging==25.0 \
protobuf==6.33.3 psutil==7.2.1 pydantic==2.12.5 \
pydantic-core==2.41.5 pyyaml==6.0.3 qwix==0.1.5 \
rich==14.2.0 simplejson==3.20.2 tensorboardx==2.6.4 \
tensorstore==0.1.80 tokamax==0.0.11 treescope==0.1.10 \
typing-inspection==0.4.2 wadler-lindig==0.1.7 zipp==3.23.0 biopython==1.83
pip install pandas==2.2.3 matplotlib==3.10.0 absl-py==2.1.0 chex==0.1.87 \
dm-haiku==0.0.13 dm-tree==0.1.8 filelock==3.16.1 \
"jax[cuda12]==0.4.34" jax-cuda12-pjrt==0.4.34 jax-triton==0.2.0 \
jaxlib==0.4.34 jaxtyping==0.2.34 jmp==0.0.4 ml-dtypes==0.5.0 \
numpy==2.1.3 nvidia-cublas-cu12==12.6.3.3 \
nvidia-cuda-cupti-cu12==12.6.80 nvidia-cuda-nvcc-cu12==12.6.77 \
nvidia-cuda-runtime-cu12==12.6.77 nvidia-cudnn-cu12==9.5.1.17 \
nvidia-cufft-cu12==11.3.0.4 nvidia-cusolver-cu12==11.7.1.2 \
nvidia-cusparse-cu12==12.5.4.2 nvidia-nccl-cu12==2.23.4 \
nvidia-nvjitlink-cu12==12.6.77 opt-einsum==3.4.0 pillow==11.0.0 \
rdkit==2024.3.5 scipy==1.14.1 tabulate==0.9.0 toolz==1.0.0 \
tqdm==4.67.0 triton==3.1.0 typeguard==2.13.3 \
typing-extensions==4.12.2 zstandard==0.23.0pip install pandas==2.2.3 matplotlib==3.10.0 absl-py==2.1.0 chex==0.1.87 \
dm-haiku==0.0.14 dm-tree==0.1.8 filelock==3.16.1 \
"jax[cuda12]==0.6.1" jax-cuda12-pjrt==0.6.1 jax-cuda12-plugin[with-cuda]==0.6.1 \
jax-triton==0.3.0 jaxlib==0.6.1 jaxtyping==0.2.34 jmp==0.0.4 ml-dtypes==0.5.0 \
numpy==2.1.3 nvidia-cublas-cu12==12.8.4.1 \
nvidia-cuda-cupti-cu12==12.9.19 nvidia-cuda-nvcc-cu12==12.9.41 \
nvidia-cuda-runtime-cu12==12.9.37 nvidia-cudnn-cu12==9.10.1.4 \
nvidia-cufft-cu12==11.4.0.6 nvidia-cusolver-cu12==11.7.4.40 \
nvidia-cusparse-cu12==12.5.9.5 nvidia-nccl-cu12==2.26.5 \
nvidia-nvjitlink-cu12==12.9.41 opt-einsum==3.4.0 pillow==11.0.0 \
rdkit==2024.3.5 scipy==1.14.1 tabulate==0.9.0 toolz==1.0.0 \
tqdm==4.67.0 triton==3.3.1 typeguard==2.13.3 \
typing-extensions==4.12.2 zstandard==0.23.0
Note: The AlphaFold 3.0.3 installation instructions have been tested on an RTX 4090.
Examples of supported graphics accelerators:
Ampere & Ada Lovelace: RTX 3060, RTX 3080, RTX 3090, RTX 4060–4090, A100, L40.
Blackwell: RTX 5080–5090, B200, GB200.Many thanks to lonelycrab888 and xryller for their contributions supporting the NVIDIA Blackwell architecture.
Define your desired installation location and clone the official repository:
# Set the desired application directory
export APPDIR="/home/user/Programs" # Replace "/home/user/Programs" with your desired path
# Create the directory and navigate to it
mkdir -p $APPDIR
cd $APPDIR
# Clone the AlphaFold 3 repository
git clone https://github.com/google-deepmind/alphafold3.git
# Define the AlphaFold 3 directory variable
export ALPHAFOLD3DIR="$APPDIR/alphafold3"
cd ${ALPHAFOLD3DIR}Important: Use
setenvinstead of export in C shell (csh/tcsh)
The fetch_databases.sh script downloads the necessary databases for AlphaFold 3. By default, it downloads them to your home directory. To change the download location to ${ALPHAFOLD3DIR}, modify the script:
# Modify the download path in the script
sed -i 's|$HOME|$ALPHAFOLD3DIR|g' fetch_databases.sh
# Make the script executable
chmod +x fetch_databases.sh
# Run the script to download the databases
./fetch_databases.shImportant: Ensure you have sufficient disk space, as the databases are quite large (~627GB).
If you use C shell (csh/tcsh), run thefetch_databases.cshscript included in the repository.
You need to request access to the AlphaFold 3 model parameters:
- Complete the official request form provided by the AlphaFold team.
- Access will be granted at Google DeepMind’s sole discretion. They aim to respond to requests within 2–3 business days. You may only use AlphaFold 3 model parameters if received directly from Google. Use is subject to these terms of use.
- Once approved, you will receive a download link for
af3.bin.zst.
Proceed to decompress and move the model files:
# Download the model parameters (replace <your_download_url>)
wget <your_download_url>
# Create the model folder
mkdir ${ALPHAFOLD3DIR}/models/
# Move the compressed model to the models directory
mv af3.bin.zst ${ALPHAFOLD3DIR}/models/
# Navigate to the models directory
cd ${ALPHAFOLD3DIR}/models/
# Decompress the model
unzstd af3.bin.zstThere might be issues with zlib not linking correctly. Use the following environment variables to resolve this:
cd ${ALPHAFOLD3DIR}
# Export paths for zlib
export CXXFLAGS="-I$(dirname $(find ${CONDA_PREFIX} -name zlib.h | head -n 1))"
export LDFLAGS="-L$(dirname $(find ${CONDA_PREFIX} -name libz.so | head -n 1)) -lz"
# Install AlphaFold 3 without additional dependencies
python -m pip install --no-deps .Compile the necessary components from the conda environment's bin directory:
cd ${CONDA_PREFIX}/bin
# Execute the build script
./build_data # ExecuteVerify that the installation was successful by displaying the help message:
cd ${ALPHAFOLD3DIR}
# Display the help message
python run_alphafold.py --helpTo run run_alphafold.py from any location, add the shebang line with the path to Python from your conda environment and make it executable:
# Add the shebang line to the script
sed -i '1s|^|#!'"$(which python)"'\n|' run_alphafold.py
# Make the script executable
chmod +x run_alphafold.pyCreate a symbolic link to the script in your conda environment's bin directory:
ln -s ${PWD}/run_alphafold.py ${CONDA_PREFIX}/bin/run_alphafold.pyIf a segmentation fault occurs when initializing pjrt_plugin, it may be due to import order conflicts between JAX and other C++/CUDA extensions (e.g. alphafold3.cpp, SciPy, RDKit).
A simple workaround is to ensure jax and jnp are imported before any other modules.
This command moves the imports to the top (before from collections.abc ...) and removes duplicate occurrences later in the file:
sed -i".bk" -e 's|from collections\.abc import Callable, Sequence|import jax\nfrom jax import numpy as jnp\n\nfrom collections.abc import Callable, Sequence|' \
-e '/^import jax$/d' \
-e '/^from jax import numpy as jnp$/d' ${PWD}/run_alphafold.pyOnce you have installed AlphaFold 3, go to your working directory and test the AlphaFold 3 run using, for example, the following input JSON file named fold_input.json:
{
"name": "2PV7",
"sequences": [
{
"protein": {
"id": ["A", "B"],
"sequence": "GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG"
}
}
],
"modelSeeds": [1],
"dialect": "alphafold3",
"version": 1
}
Create a script called AF3_run.sh with the following content to facilitate running AlphaFold 3:
#!/bin/bash
APPDIR="/home/user/Programs" # Replace with your actual path if different
ALPHAFOLD3DIR="$APPDIR/alphafold3"
#HMMER3_BINDIR="/usr/bin" # Path to HMMER binaries (**installed via OS package manager or specify your path**)
HMMER3_BINDIR="${CONDA_PREFIX}/bin/" # Path to Conda binarys (**installed via conda**)
DB_DIR="${ALPHAFOLD3DIR}/public_databases"
MODEL_DIR="${ALPHAFOLD3DIR}/models"
WORK_DIR=$(pwd)
OUTPUT_DIR="${WORK_DIR}/output/${BASE_NAME}"
LOG_FILE="${OUTPUT_DIR}/af3_run.log"
JSON_FILE=$(ls -1 *.json 2>/dev/null | head -n 1)
run_alphafold.py \
--jackhmmer_binary_path="${HMMER3_BINDIR}/jackhmmer" \
--nhmmer_binary_path="${HMMER3_BINDIR}/nhmmer" \
--hmmalign_binary_path="${HMMER3_BINDIR}/hmmalign" \
--hmmsearch_binary_path="${HMMER3_BINDIR}/hmmsearch" \
--hmmbuild_binary_path="${HMMER3_BINDIR}/hmmbuild" \
--db_dir="${DB_DIR}" \
--model_dir="${MODEL_DIR}" \
--json_path="${WORK_DIR}/${JSON_FILE}" \
--output_dir="${OUTPUT_DIR}" \
--buckets="256,512,768,1024,1280,1536,2048,2560,3072,3584,4096,4608,5120" \
2>&1 | tee -a "${LOG_FILE}"Important: Use
setin C shell (csh/tcsh) to initialize APPDIR, ALPHAFOLD3DIR,HMMER3_BINDIR, DB_DIR, MODEL_DIR, WORK_DIR, OUTPUT_DIR and LOG_FILE JSON_FILE. e.g.set APPDIR="/home/user/Programs"
Make the script executable:
chmod +x AF3_run.shNow, with your conda environment active and in the appropriate working directory, you can run AlphaFold 3 simply by executing:
./AF3_run.sh-
Use the
AF3_run.shscript anywhere on the system.cp AF3_run.sh ${CONDA_PREFIX}/bin/ chmod +x ${CONDA_PREFIX}/bin/AF3_run.sh -
Hardware Requirements: To run AlphaFold 3 on systems with limited resources, a minimum of an Amper NVIDIA GPU with 8 o 12 GB of VRAM. However, for optimal performance, it's recommended to use professional GPUs like the NVIDIA A100, H100 or high-end consumer GPUs such as the RTX 3090, 4090, or the latest 5090, as these offer superior memory and processing capabilities that significantly enhance the efficiency of running AlphaFold 3.
Verify that you have the correct versions of CUDA and NVIDIA drivers that match the installed nvidia packages.
It is recommended to install the latest graphics card driver, particularly the one provided by each distribution, especially when installing on Linux.
sudo apt-get install nvidia-driver-565
sudo rebootsudo yum install epel-release
sudo yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
sudo yum install nvidia-driver-565
sudo reboot- Disk Space: The databases (627 GB), the models (1.1 GB), and the Conda environment (6.7 GB) require significant disk space. Make sure you have at least 700 GB of free space available. For optimal performance, a dedicated 1 TB NVMe drive is recommended.
- Updates and Support: Regularly check the official AlphaFold 3 repository for updates and potential changes to dependencies.
If you encounter any issues during the installation or execution process, feel free to ask for additional assistance.
This work was depeloped by Hiiragizawa and inspired by the Alphafold3-Fedora-install repository by ullahsamee, originally developed for Fedora. We have adapted it to enhance its general applicability.