This is a Pytorch implementation of Double-head Transformer Neural Network for Molecular Property Prediction (DHTNN)
Figure 1 Overall DHTNN architectural diagram
python=3.8.10
pytorch=1.4.0
torchvision=0.5.0
chemprop=1.3.0
flask=2.0.1
hyperopt=0.2.5
matplotlib=3.5.1
numpy=1.20.3
pandas=1.2.4
pandas-flavor=0.2.0
pip=21.1.1
rdkit=2021.03.2
scikit-learn=0.24.2
scipy=1.6.3
tensorboardX=2.2
tqdm=4.61.0
typed-argument-parser=1.6.2
git+https://github.com/bp-kelley/descriptastorus
There are six datasets, which are Lipophilicity, PDBbind, PCBA, BACE, Tox21, and SIDER.
You can be get the data by uncompressing data.zip. All the data used in the experiments is in here.
chemprop_train --data_path <path> --dataset_type <type> --save_dir <dir>
where <path> is the path to a CSV file containing a dataset, <type> is one of [classification, regression] depending on the type of the dataset, and <dir> is the directory where model checkpoints will be saved.
We refer to the paper of Analyzing Learned Molecular Representations for Property Prediction. We are grateful for the previous work of swansonk14 team.