AlphaFold 3.0.0 | Research Center for Computational Science

AlphaFold3 model parameters must be requested and downloaded by the user themselves.

Webpage

https://github.com/google-deepmind/alphafold3

Version

3.0.0 (latest code on Nov 26, 2024; commit it: cdbcf41b71235)

Build Environment

NVIDIA driver 560.35.03
Miniforge3-Linux-x86_64.sh (miniforge installer)

Installation procedure

This installation procedure is taken from the notes at the time of installation. It may differ from the actual procedure. Basically, most of the actions were performed just after the release (Nov 12), but some additional actions were required due to the code updates.

hmmer-3.4

$ wget http://eddylab.org/software/hmmer/hmmer-3.4.tar.gz
$ tar zxf hmmer-3.4.tar.gz
$ ./configure --prefix /apl/alphafold/hmmer-3.4/
$ make -j8
$ make install
$ cd easel/
$ make install

python env (conda+pip)+alphafold3

$ sh Miniforge3-Linux-x86_64.sh
...
[.....] >>> /apl/alphafold/miniforge1-3.0.0
...
$ cd /apl/alphafold
$ /apl/alphafold/miniforge1-3.0.0/bin/conda shell.bash hook > af300_init.sh
$ /apl/alphafold/miniforge1-3.0.0/bin/conda shell.csh hook > af300_init.csh
$ vi af300_init.sh
(末尾に以下の行を追加)
export PATH="/apl/alphafold/hmmer-3.4/bin:$PATH"
$ vi af300_init.csh
(末尾に以下の行を追加)
setenv PATH "/apl/alphafold/hmmer-3.4/bin:$PATH"
$ . af300_init.sh
$ conda create -n af3 python=3.11
$ conda activate af3
$ sed -i -e "s/base/af3/" af300_init.sh af300_init.csh
$ conda install -c nvidia absl-py=2.1.0 \
chex=0.1.87 \
dm-tree=0.1.8 \
filelock=3.16.1 \
jaxtyping=0.2.34 \
jmp=0.0.4 \
ml_dtypes=0.5.0 \
numpy=2.1.3 \
opt_einsum=3.4.0 \
pillow=11.0.0 \
rdkit=2024.03.5 \
scipy=1.14.1 \
tabulate=0.9.0 \
toolz=1.0.0 \
tqdm=4.67.0 \
typeguard=2.13.3 \
typing-extensions=4.12.2 \
zstandard=0.23.0 \
cuda
$ pip install jax[cuda12]==0.4.34 \
jaxlib==0.4.34 \
jmp==0.0.4 \
chex==0.1.87 \
opt-einsum==3.4.0 \
dm-haiku==0.0.13 \
triton==3.1.0 \
jax-triton==0.2.0
$ cd /some/where
$ git clone https://github.com/google-deepmind/alphafold3.git
$ cd alphafold
$ git pull
$ cd ../
$ cp -r alphafold /apl/alphafold/3.0.0-20241126
$ cd /apl/alphafold
$ ln -s ./3.0.0-20241126 3.0.0
$ cd 3.0.0
$ pip install --no-deps .
$ build_data

Note: package list is not very carefully verified. Some of actions might not be necessary.
- We can skip lengthy "hmmer" executable path specifications for "run_alphafold.py" if the bin directory is involved in PATH.
- Some pip packages are necessary. Conda packages (on conda-forge) may not be enough.

DB

$ cd /apl/alphafold/3.0.0
$ python3 fetch_databases.py --download_destination=/apl/alphafold/databases3/20241112
$ cd /apl/alphafold/databases3/20241112
$ tar xf pdb_2022_09_28_mmcif_files.tar

Note: On Nov 12, we didn't need to untar "pdb_2022_09_28_mmcif_files.tar". However, expanded one was necessary on Nov 26.
- In the procedure above, "python3 fetch_databases.py" was performed on Nov 12, 2024. pdb_2022_09_28_mmcif_files.tar was expaneded on Nov 26, 2024.
- Manual untar is not necessary if the latest "fetch_databases.py" is used.
- (chmod and chown actions were omitted in the procedure above.)

For Lustre filesystem, migration of database files

$ lfs migrate -c 2 bfd-first_non_consensus_sequences.fasta
$ lfs migrate -c 10 mgy_clusters_2022_05.fa
$ lfs migrate -c 6 nt_rna_2023_02_23_clust_seq_id_90_cov_80_rep_seq.fasta
$ lfs migrate -c 18 pdb_2022_09_28_mmcif_files.tar
$ lfs migrate -c 2 rnacentral_active_seq_id_90_cov_80_linclust.fasta
$ lfs migrate -c 8 uniprot_all_2021_04.fa
$ lfs migrate -c 5 uniref90_2022_05.fa

wrapper script (run-af-300.sh)

#!/bin/bash
#
# wrapper for alphafold3 non-container version
#
# NOTE: This script is designed for AlphaFold 3.0.0.
# set default params
AFROOT=/apl/alphafold
AF3ROOT=${AFROOT}/3.0.0 # where run_alphafold.py resides
# default database path
DB_DIR=${AFROOT}/databases3/20241112
MYOPTS=""
ECHO_COMMAND=false
NO_DB_SPEC=true
DRY_RUN=false
HAS_INPUT=false
HAS_OUTPUT=false

usage() {
echo ""
echo "Usage: $0 <OPTIONS>"
echo ""
echo "Options:"
echo " -j <input json>, --json_path=<input json>"
echo " Path to input json file (single input file)."
echo " -i <input directory>, --input_dir=<input directory>"
echo " Path to directory which contains input json files (multiple input files)."
echo " -o <output directory>, --output_dir=<output directory>"
echo " Path to output directory."
echo " -m <model directory>, --model_dir=<model directory>"
echo " Path to model for inference (THIS IS NOT PROVIDED BY RCCS)."
echo " -M, --msa, --msaonly"
echo " Do only data pipeline (MSA)."
echo " -I, --inference, --inferenceonly"
echo " Skip MSA and do only inference."
echo ""
echo " Other run_alphafold.py options may also be accepted."

exit 1
}
while getopts "a:d:ehi:j:o:m:IM-:" c; do
optarg="${OPTARG}"
if [[ "$c" = - ]]; then
c="-${OPTARG%%=*}"
optarg="${OPTARG/${OPTARG%%=*}/}"
optarg="${optarg#=}"
if [[ -z "${optarg}" ]] && [[ ! "${!OPTIND}" = -* ]]; then
optarg="${!OPTIND}"
shift
fi
fi

case "-$c" in
-a|--af3root)
AF3ROOT=${optarg}
;;
-d|--db_dir)
MYOPTS="$MYOPTS --db_dir=${optarg}"
NO_DB_SPEC=false
;;
--dryrun)
ECHO_COMMAND=true
DRY_RUN=true
;;
-e|--echo)
ECHO_COMMAND=true
;;
-i|--input_dir)
MYOPTS="$MYOPTS --input_dir=${optarg}"
HAS_INPUT=true
;;
-j|--json_path)
MYOPTS="$MYOPTS --json_path=${optarg}"
HAS_INPUT=true
;;
-h|--help)
usage
;;
-m|--model_dir)
MYOPTS="$MYOPTS --model_dir=${optarg}"
;;
-o|--output_dir)
MYOPTS="$MYOPTS --output_dir=${optarg}"
HAS_OUTPUT=true
;;
-I|--inference|--inferenceonly|--norun_data_pipeline)
MYOPTS="$MYOPTS --norun_data_pipeline"
;;
-M|--msa|--msaonly|--norun_inference)
MYOPTS="$MYOPTS --norun_inference"
;;
--flash_attention_implementation)
MYOPTS="$MYOPTS --flash_attention_implementation=${optarg}"
;;
--jackhmmer_binary_path)
MYOPTS="$MYOPTS --jackhmmer_binary_path=${optarg}"
;;
--nhmmer_binary_path)
MYOPTS="$MYOPTS --nhmmer_binary_path=${optarg}"
;;
--hmmalign_binary_path)
MYOPTS="$MYOPTS --hmmalign_binary_path=${optarg}"
;;
--hmmsearch_binary_path)
MYOPTS="$MYOPTS --hmmsearch_binary_path=${optarg}"
;;
--hmmbuild_binary_path)
MYOPTS="$MYOPTS --hmmbuild_binary_path=${optarg}"
;;
--small_bfd_database_path)
MYOPTS="$MYOPTS --small_bfd_database_path=${optarg}"
;;
--mgnify_database_path)
MYOPTS="$MYOPTS --mgnify_database_path=${optarg}"
;;
--uniprot_cluster_annot_database_path)
MYOPTS="$MYOPTS --uniprot_cluster_annot_database_path=${optarg}"
;;
--uniref90_database_path)
MYOPTS="$MYOPTS --uniref90_database_path=${optarg}"
;;
--ntrna_database_path)
MYOPTS="$MYOPTS --ntrna_database_path=${optarg}"
;;
--rfam_database_path)
MYOPTS="$MYOPTS --rfam_database_path=${optarg}"
;;
--rna_central_database_path)
MYOPTS="$MYOPTS --rna_central_database_path=${optarg}"
;;
--pdb_database_path)
MYOPTS="$MYOPTS --pdb_database_path=${optarg}"
;;
--seqres_database_path)
MYOPTS="$MYOPTS --seqres_database_path=${optarg}"
;;
--jackhmmer_n_cpu)
MYOPTS="$MYOPTS --jackhmmer_n_cpu=${optarg}"
;;
--nhmmer_n_cpu)
MYOPTS="$MYOPTS --nhmmer_n_cpu=${optarg}"
;;
--jax_compilation_cache_dir)
MYOPTS="$MYOPTS --jax_compilation_cache_dir=${optarg}"
;;
esac
done
if "${NO_DB_SPEC}"; then
MYOPTS="$MYOPTS --db_dir=${DB_DIR}"
fi

if "${ECHO_COMMAND}"; then
echo "python ${AF3ROOT}/run_alphafold.py $MYOPTS"
exit 0
fi
if ! "${HAS_INPUT}" || ! "${HAS_OUTPUT}"; then
echo "ERROR: Input data (-j or -i) and output directory (-o) must be specified."
usage
fi
if ! "${DRY_RUN}"; then
python ${AF3ROOT}/run_alphafold.py $MYOPTS
fi

Sample Run (two step)

In our system, data pipeline and inference parts should be run separately.

data pipeline (MSA)

#!/bin/sh
#PBS -l select=1:ncpus=64:mpiprocs=1:ompthreads=64
#PBS -l walltime=72:00:00
if [ ! -z "${PBS_O_WORKDIR}" ]; then
cd "${PBS_O_WORKDIR}"
fi
module -s purge
. /apl/alphafold/af300_init.sh
RUNAF3=/apl/alphafold/run-af-300.sh
$RUNAF3 \
-j af3-input.json \
-o af_output \
-M

Please check this page for input json file.

Inference

#!/bin/sh
#PBS -l select=1:ncpus=16:mpiprocs=1:ompthreads=16:ngpus=1
#PBS -l walltime=72:00:00
if [ ! -z "${PBS_O_WORKDIR}" ]; then
cd "${PBS_O_WORKDIR}"
fi
module -s purge
. /apl/alphafold/af300_init.sh
RUNAF3=/apl/alphafold/run-af-300.sh
$RUNAF3 \
-j af_output/2pv7/2pv7_data.json \
-o af_output \
-m ~/AlphaFold3/models \
-I

The input json file for the inference is the output of data pipeline run.
You need to prepare model parameter file by yourselves.

View PDF