This is the official implementation of the paper "A Deep Learning Approach for Rational Ligand Generation with Property Control via Reactive Building Blocks". Additionally, we offer a user-friendly online platform to implement the functionality of DeepBlock.
Github - Docker Hub - Online Platform
git clone [email protected]:BioChemAI/DeepBlock.git
cd deepblock
conda env create -f environment.yml
conda activate deepblock_env
pip install -e .
Also, for Docker
git clone [email protected]:BioChemAI/DeepBlock.git
cd deepblock
docker build --target base -t deepblock .
docker run -it --rm deepblock
The image has been uploaded to Docker Hub.
docker pull pillarszhang/deepblock:20240801A
Quick start, --input-type
can be 'seq', 'pdb', 'url', 'pdb_fn', 'pdb_id'.
python scripts/quick_start/generate.py \
--input-data 4IWQ \
--input-type pdb_id \
--num-samples 16 \
--output-json tmp/generate_result.json
Docker.
docker run --rm \
-v "$(pwd)"/tmp:/app/tmp \
pillarszhang/deepblock:20240801A \
bash -lc "python scripts/quick_start/generate.py \
--input-data 4IWQ \
--input-type pdb_id \
--num-samples 16 \
--output-json tmp/generate_result.json"
It is recommended to use VSCode for development, as debugging configuration files are already available in .vscode
.
wget https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/releases/chembl_31/chembl_31_chemreps.txt.gz
gzip -dk chembl_31_chemreps.txt.gz
Finally, run the script
python scripts/preprocess/chembl.py \
--chembl-chemreps <PATH TO chembl_31_chemreps.txt>
CrossDocked Dataset (Index by 3D-Generative-SBDD)
Download from the compressed package we provide https://figshare.com/articles/dataset/crossdocked_pocket10_with_protein_tar_gz/25878871 (recommended). The alternative method is to obtain the files from the 3D-Generative-SBDD's index file and the raw data for the CrossDocked2020 set. The script will re-fetch the required files.
tar xzf crossdocked_pocket10_with_protein.tar.gz
The following files are required to exist:
$sbdd_dir/split_by_name.pt
$sbdd_dir/index.pkl
$sbdd_dir/1B57_HUMAN_25_300_0/5u98_D_rec_5u98_1kx_lig_tt_min_0_pocket10.pdb
$sbdd_dir/1B57_HUMAN_25_300_0/5u98_D_rec.pdb
(Recommended method)$crossdocked_dir/1B57_HUMAN_25_300_0/5u98_D_rec.pdb
(Alternative method)
Finally, run the script
python scripts/preprocess/crossdocked.py \
--sbdd-dir <PATH TO crossdocked_pocket10_with_protein> \
# --crossdocked-dir <PATH TO CrossDocked2020> # Not needed when using the recommended method
Please download the following 3 files from the PDBbind v2020
to the same directory (assuming it is $pdbbind_dir
).
- Index files of PDBbind -> PDBbind_v2020_plain_text_index.tar.gz
- Protein-ligand complexes: The general set minus refined set -> PDBbind_v2020_other_PL.tar.gz
- Protein-ligand complexes: The refined set -> PDBbind_v2020_refined.tar.gz
To extract the files, first navigate to the $pdbbind_dir
directory and then use the following command.
tarballs=("PDBbind_v2020_plain_text_index.tar.gz" "PDBbind_v2020_refined.tar.gz" "PDBbind_v2020_other_PL.tar.gz")
for tarball in "${tarballs[@]}"
do
dirname=${tarball%%.*}
mkdir -p "$dirname" && pv -N "Extracting $tarball" "$tarball" | tar xzf - -C "$dirname"
done
The following files are required to exist:
$pdbbind_dir/PDBbind_v2020_plain_text_index/index/INDEX_general_PL.2020
$pdbbind_dir/PDBbind_v2020_refined/refined-set/1a1e/1a1e_ligand.sdf
$pdbbind_dir/PDBbind_v2020_other_PL/v2020-other-PL/1a0q/1a0q_protein.pdb
Finally, run the script
python scripts/preprocess/pdbbind.py \
--pdbbind-dir <PATH TO PDBBind>
# Filter test set
python scripts/preprocess/pdbbind_pick_set.py
python scripts/preprocess/merge_vocabs.py \
--includes chembl crossdocked
python scripts/cvae_complex/train.py \
--include chembl \
--device cuda:0 \
--config scripts/cvae_complex/frag_pretrain_config.yaml \
--vocab-fn "saved/preprocess/merge_vocabs/chembl,crossdocked&frag_vocab.json" \
--no-valid-prior
Replace 20230303_191022_be9e
with pretrain ID.
python scripts/cvae_complex/train.py \
--include crossdocked \
--device cuda:0 \
--config scripts/cvae_complex/complex_config.yaml \
--base-train-id 20230303_191022_be9e
The checkpoint trained on ChEMBL and CrossDocked has been provided in the repository at
deepblock/public/saved/cvae_complex/20230305_163841_cee4
. By executingcp -r deepblock/public/saved .
, you can directly continue with the following commands.
Replace 20230305_163841_cee4
with train ID.
python scripts/cvae_complex/sample.py \
--include crossdocked \
--device cuda:0 \
--base-train-id 20230305_163841_cee4 \
--num-samples 100 \
--validate-mol \
--embed-mol \
--unique-mol
python scripts/cvae_complex/optimize.py \
--device cpu \
--base-train-id 20230305_163841_cee4 \
--num-samples 5000 \
--complex-id F16P1_HUMAN_1_338_0/3kc1_A_rec_3kc1_2t6_lig_tt_min_0
Train a drug toxicity decision tree predictor based on molecular fingerprints. The script will automatically download the dataset from TOXRIC and use TPOT for automatic parameter tuning.
The trained checkpoint has already been provided in the repository at
deepblock/public/saved/regress_tox/20230402_135859_c91a
.
python scripts/regress_tox/train.py
python scripts/cvae_complex/sample_sa.py \
--device cpu \
--base-train-id 20230305_163841_cee4 \
--num-samples 100 \
--num-steps 50 \
--complex-id F16P1_HUMAN_1_338_0/3kc1_A_rec_3kc1_2t6_lig_tt_min_0
Automatically download and configure tools such as ADFR Suite and QuickVina2, and check the docking toolchain. They will be deployed to the work/docking_toolbox
directory in the working directory.
python scripts/evaluate/init_docking_toolbox.py
The prepare_batch_docking.py
script will retrieve previously sampled molecules from the --base-train-id
folder, then deduplicate and prepare pdbqt files. It will construct receptor-ligand pairs, generate hashes, compress and create a docking task package, as well as generate a lookup table for the sampled results and docking task hashes.
python scripts/evaluate/prepare_batch_docking.py \
--include crossdocked \
--sbdd-dir ~/dataset/crossdocked_pocket10_with_protein \
--base-train-id 20230305_163841_cee4 \
--suffix _100veu \
--n-jobs 8 \
--dock-backend qvina2
Next, the docking task package can be transferred to other computers or shared computing platforms to execute batch docking tasks with any number of processes.
python scripts/evaluate/run_batch_docking.py \
--dock-backend qvina2 \
--input saved/cvae_complex/20230305_163841_cee4/evalute/batch_docking_input_100veu_qvina2.7z \
--n-procs 40
After this script finishes, it will automatically package the docking results. Additionally, it will copy the JSON file containing the hash-score to the same folder as the docking task package. You can then copy it back to your computer.
python scripts/evaluate/compute_qedsa.py \
--base-train-id 20230305_163841_cee4 \
--suffix _100veu
python scripts/evaluate/compute_dist.py \
--base-train-id 20230305_163841_cee4 \
--suffix _100veu
The script will compile the above-generated metrics to produce the final mean and variance.
python scripts/evaluate/summary.py \
--base-train-id 20230305_163841_cee4 \
--suffix _100veu \
--docking-suffix _100veu_qvina2