liGAN is a PyTorch project for structure-based drug discovery with deep generative models of atomic density grids.

• CVAE journal paper (under review)

• CVAE NeurIPS 2020 workshop paper (15 minute talk)

• VAE NeurIPS 2020 workshop paper (2 minute talk)

• numpy
• pandas
• scikit-image
• torch
• openbabel
• rdkit
• molgrid
• gnina

NOTE: Please be aware that the current version of molgrid provided through pip/conda is incompatible with conda openbabel, and you will likely get segmentation faults if you install them both through conda.

A molgrid conda build recipe is in the works (see https://github.com/mattragoza/conda-molgrid), but for now you can use this environment to build libmolgrid from source.

To generate molecules, you must first download the pretrained model weights:

sh download_weights.sh

Then just run the generate.py script with the default configuration file:

python3 generate.py config/generate.config

To train a model from scratch, you must first download the full Crossdocked2020 data set:

sh download_data.sh

More info about this data set can be found here.

Then you can run the train.py script with the default configuration file:

python3 train.py config/train.config

Coming soon!