Sampling a Trillion-Sized GDB-20 Database of Drug-Like Molecules by Generative Artificial Intelligence
Thank you for your interest in this repository, which complements the publication "Sampling a Trillion-Sized GDB-20 Database of Drug-Like Molecules by Generative Artificial Intelligence".
For a step-by-step map from the manuscript workflow to the repository files, software environments, graph-selection procedure, and executable scripts, see REPRODUCIBILITY.rst.
Zenodo Part 1:
# See https://zenodo.org/records/17368725.Zenodo Part 2:
# See https://zenodo.org/records/17375415.GDB-ML/
├── src/
| └── gdb_ml/
| ├── chem_utils.py
| ├── data_processor.py
| ├── graph_mapping.py
| └── properties_calculator.py
├── transformer/
| ├── pipeline.ipynb
| ├── preprocess.py
| ├── train.py
| ├── translate.py
| ├── gdb20_data
| └── gdb20_model
└── generative_models/
├── create_randomized_smiles.py
├── create_model.py
├── train_model.py
├── sample_from_model.py
├── calculate_nlls.py
├── gdb20_data
└── gdb20_models
(1) Create a Conda environment with the provided `.yaml` file and activate it:
conda env create -f environment-gdb20.yaml
conda activate gdb20(2) Follow the pipeline and tokenize the SMILES:
# See transformer/pipeline.ipynb(3) OpenNMT-py Installation:
conda env create -f transformer/environment-opennmt.yaml
conda activate opennmt
git clone https://github.com/reymond-group/OpenNMT-py.git
cd OpenNMT-py
git checkout Enzymatic_Transformer
pip install -e .
cd ..(4) Preprocess the data:
# Concatenate split files into manuscript-style filenames
cat transformer/gdb20_data/shuffled_train_keys_part_*_canonical_concatenated_tokenized.txt \
> transformer/gdb20_data/src_train.txt
cat transformer/gdb20_data/shuffled_train_values_part_*_canonical_concatenated_tokenized.txt \
> transformer/gdb20_data/tgt_train.txt
cat transformer/gdb20_data/shuffled_val_keys_part_*_canonical_concatenated_tokenized.txt \
> transformer/gdb20_data/src_val.txt
cat transformer/gdb20_data/shuffled_val_values_part_*_canonical_concatenated_tokenized.txt \
> transformer/gdb20_data/tgt_val.txt
cat \
"transformer/gdb20_model/test36_model_step_55000_part_aa.pt" \
"transformer/gdb20_model/test36_model_step_55000_part_ab.pt" \
"transformer/gdb20_model/test36_model_step_55000_part_ac.pt" \
> "transformer/gdb20_model/test36_model_step_55000.full.pt"
# Define variables
data_dir="gdb20_data"
dataset="test36"
experiment="exp36"
batchsize=6144
dropout=0.1
rnnsize=384
wordvecsize=384
learnrate=2
layers=4
heads=8
MODEL_PATH="gdb20_model/test36_model_step_55000.full.pt"
SRC_FILE="gdb20_data/src_val.txt"
OUTPUT_FILE="experiments/test36_model_step_55000_predictions.txt"
# Run the remaining transformer commands from this directory.
cd transformer
mkdir -p data/voc_${experiment}
# Run preprocessing
python preprocess.py \
-train_src "${data_dir}/src_train.txt" \
-train_tgt "${data_dir}/tgt_train.txt" \
-valid_src "${data_dir}/src_val.txt" \
-valid_tgt "${data_dir}/tgt_val.txt" \
-save_data data/voc_${experiment}/Preprocessed \
-src_seq_length 3000 -tgt_seq_length 3000 \
-src_vocab_size 3000 -tgt_vocab_size 3000 \
-share_vocab -lower(5) Train the transformer model:
# Remove the line "-gpu_ranks 0 \"" when training on CPU
python train.py \
-data data/voc_${experiment}/Preprocessed \
-save_model experiments/checkpoints/${experiment}/${dataset}_model \
-seed 42 \
-save_checkpoint_steps 500 \
-keep_checkpoint 50 \
-train_steps 500000 \
-param_init 0 \
-param_init_glorot \
-max_generator_batches 32 \
-batch_size ${batchsize} \
-batch_type tokens \
-normalization tokens \
-max_grad_norm 0 \
-accum_count 4 \
-optim adam \
-adam_beta1 0.9 \
-adam_beta2 0.998 \
-decay_method noam \
-warmup_steps 8000 \
-learning_rate ${learnrate} \
-label_smoothing 0.0 \
-layers 4 \
-rnn_size ${rnnsize} \
-word_vec_size ${wordvecsize} \
-encoder_type transformer \
-decoder_type transformer \
-dropout ${dropout} \
-position_encoding \
-global_attention general \
-global_attention_function softmax \
-self_attn_type scaled-dot \
-heads 8 \
-transformer_ff 2048 \
-valid_steps 500 \
-valid_batch_size 4 \
-report_every 500 \
-log_file data/Training_LOG_${experiment}.txt \
-early_stopping 10 \
-early_stopping_criteria accuracy \
-world_size 1 \
-gpu_ranks 0 \
-tensorboard \
-tensorboard_log_dir experiments/Tensorboard/${experiment}/(6) Molecular generation:
python translate.py \
-model "$MODEL_PATH" \
-src "$SRC_FILE" \
-output "$OUTPUT_FILE" \
-batch_size 64 \
-replace_unk \
-max_length 1000 \
-log_probs \
-beam_size 300 \
-n_best 300PySpark requires Java. Please install a JDK, e.g. JDK 11 or 17, and make sure
JAVA_HOME is set before running scripts that use PySpark.
Steps 5 and 6 require an NVIDIA GPU with CUDA. If using a GPU, replace
pytorch=2.7.0 with pytorch-gpu=2.7.0 in environment-gdb20.yaml before creating the environment.
(1) Activate the Conda environment:
conda activate gdb20
# While gdb20 is active, install Java:
conda install -c conda-forge openjdk=17(2) Create a working directory:
# Run the remaining RNN commands from this directory.
cd generative_models
mkdir -p node18_randomized/models(3) Create random SMILES:
./create_randomized_smiles.py -i gdb20_data/1M_node18_train.txt -o node18_randomized/training -n 100
./create_randomized_smiles.py -i gdb20_data/1M_node18_validation.txt -o node18_randomized/validation -n 100(4) Create a blank model file:
./create_model.py -i node18_randomized/training/001.smi -o node18_randomized/models/model.empty(5) Train the generative model with specified parameters:
./train_model.py \
-i node18_randomized/models/model.empty \
-o node18_randomized/models/model.trained \
-s node18_randomized/training \
-e 100 --lrm ada \
--csl node18_randomized/tensorboard \
--csv node18_randomized/validation \
--csn 75000(6) Sample an already trained model for a given number of SMILES (also retrieves log-likelihoods):
# To use the bundled pretrained model instead, replace the -m path below with:
# gdb20_models/model.trained.node18
./sample_from_model.py \
-m node18_randomized/models/model.trained.100 \
-n 1000000 \
--with-nll \
-o output.txt- If you reuse this code please also cite the underlying code frameworks: "OpenNMT technical report" and "Enzymatic_Transformer".
- If you reuse this code please also cite the underlying code framework: "reinvent-randomized".
- Free software: MIT license
This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.