Welcome to the code accompanying the paper "Constrained Bayesian Optimisation for Automatic Chemical Design using Variational Autoencoders"

https://pubs.rsc.org/en/content/articlehtml/2019/sc/c9sc04026a

The code is based heavily on the implementation of the Aspuru-Guzik group:

https://github.com/aspuru-guzik-group/chemical_vae

Append the package directory location to your PYTHONPATH e.g. by editing the .bashrc file as follows:

vim ~/.bashrc

and adding

export PYTHONPATH

source ~/.bashrc

It is recommended that you install dependencies within a virtual environment. For example, using conda you would run, from the Constrained_BO_package directory, the commands:

conda config --add channels conda-forge

(to add conda-forge to existing channels)

source activate env_name

conda install rdkit==2017.09.3

cd Theano-master

python setup.py install

cd ..

conda install numpy==1.13.0

pip install git+https://github.com/rgbombarelli/keras.git#egg=Keras

pip install git+https://github.com/rgbombarelli/seya.git#egg=seya

pip install git+https://github.com/HIPS/autograd.git#egg=autograd

The scripts

generate_latent_features_and_targets_example.py generate_qed_features_and_targets.py generate_solo_qed.py

must be run first in order to create the features and targets for molecule generation.

1. Branin_Hoo

Constrained Bayesian Optimisation on the toy Branin-Hoo function.

2. Chemical_Design

The Unconstrained directory contains scripts that generate molecules using unconstrained Bayesian Optimisation. The Constrained directory contains scripts that generate molecules using constrained Bayesian Optimisation.

Within these directories there are 3 scripts optimising the following objectives:

a) bo_gp.py -> logP + SA + ring-penalty b) bo_gp_qed -> QED + SA + ring-penalty c) bo_gp_solo_qed -> QED

The Initialisation directory contains code to generate training data for the binary classification neural network in the scripts Pos_Gen.py and Neg_Gen.py. These scripts inteface with the make_training_data.py script in order to create the data.

Sample Bibtex is given below:

@article{griffiths2020constrained,
  title={Constrained Bayesian optimization for automatic chemical design using variational autoencoders},
  author={Griffiths, Ryan-Rhys and Hern{\'a}ndez-Lobato, Jos{\'e} Miguel},
  journal={Chemical Science},
  year={2020},
  publisher={Royal Society of Chemistry}
}