AlphaFold2 without Docker

NOTES.md

Warning: this is my personal messy ad-hoc note. Not intended for production installation. Instruction by @kalinialab is probably better.

My environment

• Cent OS 7.0
• CUDA 11.2
• Without root accoumt

Directories

AF_ROOT=/path/to/somewhere
cd $AF_ROOT
mkdir tools
TOOLS=$AF_ROOT/tools
git clone https://github.com/deepmind/alphafold.git

Download database

First, manually install aria2c to somewhere in PATH.

Fix bugs in the download scripts.

--- a/scripts/download_mgnify.sh
+++ b/scripts/download_mgnify.sh
@@ -35,7 +35,7 @@ SOURCE_URL="ftp://ftp.ebi.ac.uk/pub/databases/metagenomics/peptide_database/2018
 aria2c "${SOURCE_URL}" --dir="${ROOT_DIR}"
 pushd "${ROOT_DIR}"
-gunzip "${ROOT_DIR}/${BASENAME}"
+gunzip "${BASENAME}"
 popd
index 061818c..36e1ef5 100755
--- a/scripts/download_uniref90.sh
+++ b/scripts/download_uniref90.sh
@@ -37,5 +37,5 @@ BASENAME=$(basename "${SOURCE_URL}")
 aria2c "${SOURCE_URL}" --dir="${ROOT_DIR}"
 pushd "${ROOT_DIR}"
-gunzip "${ROOT_DIR}/${BASENAME}"
+gunzip "${BASENAME}"
 popd

Also add -x 8 to aria2 if you want parallel download.

Run download scripts.

mkdir DBs
$AF_ROOT/alphafold/scripts/download_bfd.sh DBs &
$AF_ROOT/alphafold/scripts/download_pdb70.sh DBs &
$AF_ROOT/alphafold/scripts/download_uniclust30.sh DBs &
$AF_ROOT/alphafold/scripts/download_alphafold_params.sh DBs &
$AF_ROOT/alphafold/scripts/download_mgnify.sh DBs &
$AF_ROOT/alphafold/scripts/download_pdb_mmcif.sh DBs &
$AF_ROOT/alphafold/scripts/download_uniref90.sh DBs &

wget https://git.scicore.unibas.ch/schwede/openstructure/-/raw/7102c63615b64735c4941278d92b554ec94415f8/modules/mol/alg/src/stereo_chemical_props.txt
mv stereo_chemical_props.txt $AF_ROOT/alphafold/alphafold/common

Install bioinformatics tools

cd $TOOLS
git clone --branch v3.3.0 https://github.com/soedinglab/hh-suite.git hh-suite-3.3.0
cd hh-suite-3.3.0
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=$TOOLS/hh-suite ..
make -j12
make install

cd $TOOLS
wget http://eddylab.org/software/hmmer/hmmer-3.3.2.tar.gz
tar xzvf hmmer-3.3.2.tar.gz 
cd hmmer-3.3.2
./configure --prefix=$TOOLS/hmmer
make -j12
make install

cd $TOOLS
wget https://github.com/TimoLassmann/kalign/archive/refs/tags/v3.3.1.tar.gz
tar xzvf v3.3.1.tar.gz
cd kalign-3.3.1/
./autogen.sh
./configure --prefix=$TOOLS/kalign # needs GCC >= 4.9; I used 6.2.0
make -j12
make install

Create conda environment and install Python modules

warning: after pip install, conda becomes unhappy with module version conflicts and won't let you install more packages.

conda create -p $TOOLS/alphafold-conda python=3.8
conda activate $TOOLS/alphafold-conda
conda install -c conda-forge openmm=7.5.1 cudatoolkit==11.2 pdbfixer pip

pip3 install --upgrade pip
vim $AF_ROOT/alphafold/requirements.txt # comment out docker
pip3 install -r ../alphafold/requirements.txt
# My system has CUDA 11.2 but JAX has only up to 11.1
pip3 install --upgrade jax jaxlib==0.1.69+cuda111 -f https://storage.googleapis.com/jax-releases/jax_releases.html
 
cd $TOOLS/alphafold-conda/lib/python3.8/site-packages
patch -p0 < $AF_ROOT/alphafold/docker/openmm.patch
 
# Download cuDNN 8.2.1 and put lib*.so* into LD_LIBRARY_PATH
mkdir $TOOLS/cudnn
cp /path/to/cudnn/lib/* $TOOLS/cudnn

Execute

bash run.sh --fasta_paths=/absolute/path/to/sequence.fasta # You MUST use an absolute path!

A hack to use already calculated MSA

This allows you to skip slow, non-GPU accelerated MSA calculation when the GPU part failed. You can also use this to run MSA on a CPU-only node, kill the job (or add exit to the code) and then continue on a GPU node.

--- a/run_alphafold.py
+++ b/run_alphafold.py
@@ -112,15 +112,20 @@ def predict_structure(
 
   # Get features.
   t_0 = time.time()
-  feature_dict = data_pipeline.process(
-      input_fasta_path=fasta_path,
-      msa_output_dir=msa_output_dir)
-  timings['features'] = time.time() - t_0
-
-  # Write out features as a pickled dictionary.
   features_output_path = os.path.join(output_dir, 'features.pkl')
-  with open(features_output_path, 'wb') as f:
-    pickle.dump(feature_dict, f, protocol=4)
+  if os.path.exists(features_output_path):
+    feature_dict = pickle.load(open(features_output_path, 'rb'))
+
+  else:
+    feature_dict = data_pipeline.process(
+        input_fasta_path=fasta_path,
+        msa_output_dir=msa_output_dir)
+        msa_output_dir=msa_output_dir)
+  
+    # Write out features as a pickled dictionary.
+    with open(features_output_path, 'wb') as f:
+      pickle.dump(feature_dict, f, protocol=4)
+
+  timings['features'] = time.time() - t_0
 
   relaxed_pdbs = {}
   plddts = {}

A hack to run AlphaFold2 from arbitrary directories

This allows you to run AlphaFold2 from arbitrary directories, without changing the current directory to the program directory.

--- a/alphafold/common/residue_constants.py
+++ b/alphafold/common/residue_constants.py
@@ -402,8 +402,9 @@ def load_stereo_chemical_props() -> Tuple[Mapping[str, List[Bond]],
     residue_virtual_bonds: dict that maps resname --> list of Bond tuples
     residue_bond_angles: dict that maps resname --> list of BondAngle tuples
   """
-  stereo_chemical_props_path = (
-      'alphafold/common/stereo_chemical_props.txt')
+  import os.path
+  stereo_chemical_props_path = (os.path.join(os.path.dirname(__file__ ),
+      'stereo_chemical_props.txt'))
   with open(stereo_chemical_props_path, 'rt') as f:
     stereo_chemical_props = f.read()
   lines_iter = iter(stereo_chemical_props.splitlines())

run.sh

#!/bin/bash

OUTPUT=/work/AlphaFold/output

AF_ROOT=/path/to/prog/AlphaFold
TOOLS=$AF_ROOT/tools
DOWNLOAD_DIR=$AF_ROOT/DBs

source conda activate $TOOLS/alphafold-conda
export LD_LIBRARY_PATH=$TOOLS/cudnn:$LD_LIBRARY_PATH
export TF_FORCE_UNIFIED_MEMORY=1
export XLA_PYTHON_CLIENT_MEM_FRACTION=4.0
export CUDA_VISIBLE_DEVICES=0,1 # but apparently only one GPU is used...

cd $AF_ROOT/alphafold # Probably I should use PYTHONPATH...
python3.8 run_alphafold.py \
--data_dir=$DOWNLOAD_DIR \
--uniref90_database_path=$DOWNLOAD_DIR/uniref90/uniref90.fasta \
--mgnify_database_path=$DOWNLOAD_DIR/mgnify/mgy_clusters.fa \
--bfd_database_path=$DOWNLOAD_DIR/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
--uniclust30_database_path=$DOWNLOAD_DIR/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
--pdb70_database_path=$DOWNLOAD_DIR/pdb70/pdb70 \
--template_mmcif_dir=$DOWNLOAD_DIR/pdb_mmcif/mmcif_files \
--obsolete_pdbs_path=$DOWNLOAD_DIR/pdb_mmcif/obsolete.dat \
--output_dir=$OUTPUT \
--preset=casp14 \
--benchmark=False \
--model_names=model_1,model_2,model_3,model_4,model_5 \
--jackhmmer_binary_path=$TOOLS/hmmer/bin/jackhmmer \
--hhblits_binary_path=$TOOLS/hhsuite/bin/hhblits \
--hhsearch_binary_path=$TOOLS/hhsuite/bin/hhsearch \
--kalign_binary_path=$TOOLS/kalign/bin/kalign \
--max_template_date=2020-1-1 \
$@

On installation of kalign, ./autogen.sh is required before ./configure

cd $TOOLS
wget https://github.com/TimoLassmann/kalign/archive/refs/tags/v3.3.1.tar.gz
tar xzvf v3.3.1.tar.gz
cd kalign-3.3.1/
+ ./autogen.sh
./configure --prefix=$TOOLS/kalign # needs GCC >= 4.9; I used 6.2.0
make -j12
make install

Also, libtoolize package is required for ./autogen.sh. Type yum -y install libtool to install it for CentOS 7.

You are right. Thanks for pointing this out.

How to modify the code to achieve "use this to run MSA on a CPU-only node, kill the job (or add exit to the code) and then continue on a GPU node.” thank you

@chen-15969746967 Just apply the patch above.
That being said, this was written nearly four years ago and probably the latest AlphaFold has changed since then.
As I hardly use AlphaFold these days, I cannot help you.