TensorFlow implementation of the models described in the paper Molecular Geometry Prediction using a Deep Generative Graph Neural Network.

We present code for training deep generative models of molecular geometry (conformations), as well as preprocessed datasets and pretrained models.

• Python 3.6
• TensorFlow 1.2
• RDKit 2018.09.1
• tensorboardX

• CUDA (we recommend using the latest version. The version 9.0 was used in all our experiments.)

Note: Due to licensing issues, we can't release preprocessed CSD dataset. However, if you have a license to use CSD dataset please email us and we will send you the preprocessed dataset.

python PredX_train.py --data QM9 --mpnn-steps 3

python PredX_train.py --data COD --mpnn-steps 5

python PredX_train.py --data QM9 --loaddir qm9_model/neuralnet_model_best.ckpt --test --mpnn_steps 3

python PredX_train.py --data COD --loaddir cod_model/neuralnet_model_best.ckpt --test --mpnn_steps 5

python baseline.py --data QM9 --num-total-samples 100 --num-parallel-samples 10 --num-threads 10

python baseline_nn.py --data QM9 --nn-path /path/to/qm9_cvgae_confs Notes: --nn-path points to the folder containing saved conformations generated by CVGAE. These saved conformations by CVGAE can be obtained by adding --savepermol argument during loading/generation stage Example (QM9): python PredX_train.py --data QM9 --loaddir qm9_model/neuralnet_model_best.ckpt --savepermol --test --mpnn_steps 3

Instead of saving conformations by CVGAE and loading them separately, you can also run CVGAE + MMFF all together by adding --useFF argument during loading/generation stage Example (QM9): python PredX_train.py --data QM9 --loaddir qm9_model/neuralnet_model_best.ckpt --useFF --test --mpnn_steps 3

QM9_featurize.py, QM9_sdf_to_p.py, COD_featurize.py, COD_sdf_to_p.py, CSD_featurize.py, CSD_sdf_to_p.py are scripts for preprocessing QM9, COD, CSD datasets respectively

If you find the resources in this repository useful, please consider citing:

@article{Mansimov:19,
  author    = {Elman Mansimov and Omar Mahmood and Seokho Kang and Kyunghyun Cho},
  title     = {Molecular Geometry Prediction using a Deep Generative Graph Neural Network},
  year      = {2019},
  journal   = {arXiv preprint arXiv:1904.00314},
}