This repository is the official Tensorflow implementation of "Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation".

Jiaxuan You*, Bowen Liu*, Rex Ying, Vijay Pande, Jure Leskovec, Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation

• Install rdkit, please refer to the offical website for further details, using anaconda is recommended:

conda create -c rdkit -n my-rdkit-env rdkit

• Install mpi4py, networkx:

conda install mpi4py
pip install networkx=1.11

• Install OpenAI baseline dependencies:

cd rl-baselines
pip install -e .

• Install customized molecule gym environment:

cd gym-molecule
pip install -e.

There are 4 important files:

• run_molecule.py is the main code for running the program. You may tune all kinds of hyper-parameters there.
• The molecule environment code is in gym-molecule/gym_molecule/envs/molecule.py.
• RL related code is in rl-baselines/baselines/ppo1 folder: gcn_policy.py is the GCN policy network; pposgd_simple_gcn.py is the PPO algorithm specifically tuned for GCN policy.

• single process run

python run_molecule.py

• mutiple processes run

mpirun -np 8 python run_molecule.py 2>/dev/null

2>/dev/null will hide the warning info provided by rdkit package.

We highly recommend using tensorboard to monitor the training process. To do this, you may run

tensorboard --logdir runs

All the generated molecules along the training process are stored in the molecule_gen folder, each run configuration is stored in a different csv file. Molecules are stored using SMILES strings, along with the desired properties scores.