ifyoungnet / SuperAtomicCharge

Out-of-the-box Deep Learning Prediction of Atomic Partial Charges by Graph Representation and Transfer Learning

SuperAtomicCharge

Out-of-the-box Deep Learning Prediction of Atomic Partial Charges by Graph Representation and Transfer Learning. This source code was tested sucessfully on the basic environment with conda==4.5.4 and cuda==11.0

Conda Environment Reproduce

Two methods were provided for reproducing the conda environment used in this study

• create environment using file packaged by conda-pack

  Download the packaged file dgl430_v1.tar.gz and following commands can be used to reproduce the conda environment:

  mkdir /opt/conda_env/dgl430_v1
  tar -xvf dgl430_v1.tar.gz -C /opt/conda_env/dgl430_v1
  source activate /opt/conda_env/dgl430_v1
  conda-unpack

• create environment using files provided in ./envs directory

  The following commands can be used to reproduce the conda environment:

  conda create --prefix=/opt/conda_env/dgl430_v1 --file conda_packages.txt
  source activate /opt/conda_env/dgl430_v1
  pip install torch==1.3.1+cu92 torchvision==0.4.2+cu92 -f https://download.pytorch.org/whl/torch_stable.html
  pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
  pip install -r pip_packages.txt

Usage

Users can directly use our well-trained model (depoisted in ./model_save/ directory) to predicted the corresponding atomic partial charges (DDEC4, DDEC78 and RESP). Because this method was based on the 3D molecular structures. Therefore, the actual use of this method should be conducted on optimized molecules containing 3D coordinates, such as molecules optimized by MMFF force field, PM7 method and so on, and the inputs should be a sdf file containing multiple molecules with 3D coordinates. The input example was deposited in ./inputs/3cl-min.sdf or ./inputs/casp8-min.sdf. For users who want to train their own model using new datasets, we also show a model training example in the following. The corresponding training data was deposited in ./training_data. The label of training data can be assessed using the script ./scripts/get_sdf_charge.py. In addtion, we also provided a web server for directly predicting the atomic partial charges in the deepchargepredictor server

• step 1: Clone the Repository

git clone https://github.com/zjujdj/SuperAtomicCharge.git

• step 2: Construction of Conda Environment

# method1 in 'Conda Environment Reproduce' section
mkdir /opt/conda_env/dgl430_v1
tar -xvf dgl430_v1_.tar.gz -C /opt/conda_env/dgl430_v1
source activate /opt/conda_env/dgl430_v1
conda-unpack

# method2 in 'Conda Environment Reproduce' section
cd ./SuperAtomicCharge/envs
conda create --prefix=/opt/conda_env/dgl430_v1 --file conda_packages.txt
source activate /opt/conda_env/dgl430_v1
pip install torch==1.3.1+cu92 torchvision==0.4.2+cu92 -f https://download.pytorch.org/whl/torch_stable.html
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
pip install -r pip_packages.txt

• step 3: Charge Prediction

cd ./SuperAtomicCharge/scripts
nohup python3 -u model_prediction_linux.py --job_of_name=hello_charge --type_of_charge=e4 --input_file=3cl-min.sdf 
--correct_charge --device=cpu > ../outputs/prediction.log 2>&1 &

# model_prediction_linux.py use help
python3 model_prediction_linux.py -h

• step 4: Model Training Example

cd ./SuperAtomicCharge/scripts
nohup python3 -u model_train.py --gpuid=0 --lr=0.0001 --epochs=5 --batch_size=20 --tolerance=0 --patience=3 --l2=0.000001 
--repetitions=2 --type_of_charge=e4 --num_process=4 --bin_data_file=data_e4.bin > ../outputs/training.log 2>&1 &

# model_prediction_linux.py use help
python model_train.py -h

Acknowledgement

Some scripts were based on the dgl project. We'd like to show our sincere thanks to them.