Mol-CycleGAN - a generative model for molecular optimization

Official implementation of Mol-CycleGAN for molecular optimization.

Keras CycleGan implementation is based on [tjwei/GANotebooks].

Requirements

We highly recommend to use conda for package management -- the environment.yml file is provided.

The environment can be created by running:

conda env create -f environment.yml

We use Junction Tree Variational Autoencoder implementation as a submodule in Mol-CycleGAN code. After cloning this repo, the following script should be executed before running the code

./scripts/init_repo.sh 

Datasets

We provide the user with all datasets needed to reproduce the aromatic rings experiments.

Downloading all the input data (ZINC 250k dataset and related JT-VAE encodings) can be performed by running:

./scripts/download_input_data.sh

Downloading all the data from aromatic rings experiments (train / test splits of datasets, molecules returned by Mol-CycleGAN and related SMILES) can be performed by running:

./scripts/download_ar_data.sh

Basic use

This code is an implementation of CycleGan for molecular optimization.

Training of the model can be performed by running:

python train.py

with specified training parameters.

After the model is trained and the test set translation is generated, for decoding the molecules the JT-VAE code should be used. This can be performed by running:

python decode.py

with specified decoding parameters.

Experiments

We provide all the data and code needed to reproduce the Aromatic rings experiment.

1. In data/input_data/aromatic_rings/datasets_generator_aromatic_rings.ipynb one can find the data factory - the code that is needed to create train and test sets used in the experiment.

2. Training of the model can be performed by running ./scripts/run_aromatic_rings_training.sh. It calls the train.py function with base parameters, which are set to process the aromatic rings data.

3. Decoding the molecules can be performed by running ./scripts/run_aromatic_rings_decoding.sh. It calls the decode.py function with base parameters, which are set to process the aromatic rings data.

4. The analysis of the output is provided in the notebook experiments/aromatic_rings.ipynb.

Disclaimer

The code for Mol-Cycle-Gan was natively written in Python3, however, the JT-VAE package is written in Python2. To ensure the ease of use, we used downgraded versions of packages, so that the entire experiment can be run in a single environment. Since many of those packages are outdated, we strongly recommend using the environment.yml file provided to construct the working environment.