Home

Awesome

chemical VAE

This repository contains the framework and code for constructing a variational autoencoder (VAE) for use with molecular SMILES, as described in doi:10.1021/acscentsci.7b00572, with preprint at https://arxiv.org/pdf/1610.02415.pdf.

In short, molecular SMILES are encoded into a code vector representation, and can be decoded from the code representation back to molecular SMILES. The autoencoder may also be jointly trained with property prediction to help shape the latent space. The new latent space can then be optimized upon to find the molecules with the most optimized properties of interest.

In our example, we perform encoding/decoding with the ZINC dataset, and shape the latent space on prediction on logP, QED, and SAS properties.

Upcoming updates:

Questions, problems?

Make a github issue :smile:. Please be as clear and descriptive as possible.

How to install

Requirements:

An Anaconda python environment is recommend. Check the environment.yml file, but primarily:

Jupyter notebook is required to run the ipynb examples. Make sure that the Keras backend is set to use Tensorflow

via Anaconda (recommended way)

Create a conda enviroment:

conda env create -f environment.yml
source activate chemvae
python setup.py install

via pip

Assuming you have all the requirements:

pip install git+https://github.com/aspuru-guzik-group/chemical_vae.git

Example: ZINC dataset

This repository contains an example of how to run the autoencoder on the zinc dataset.

First, take a look at the zinc directory. Parameters are set in the following jsons

For a full description of all the parameters, see hyperparameters.py ; parameters set in exp.json will overwrite parameters in hyperparameters.py, and parameters set in params.json will overwrite parameters in both exp.json and hyperparameters.py

Once you have set the parameters, run the autoencoder using the command from directory with exp.json:

python -m chemvae.train_vae

(Make sure you copy examples directories to not overwrite the trained weights (*.h5))

Components

train_vae.py : main script for training variational autoencoder Accepts arguments -d ... Example of how to run (with example directory here)

Authors:

This software is written by Jennifer Wei, Benjamin Sanchez-Lengeling, Dennis Sheberla, Rafael Gomez-Bomberelli, and Alan Aspuru-Guzik (alan@aspuru.com). It is based on the work published in https://arxiv.org/pdf/1610.02415.pdf by

Feel free to reach out to us with any questions!

Funding acknowledgements

"This work was supported by the Computational Chemical Sciences Program funded by the U.S.Department of Energy, Office of Science, Basic Energy Sciences, under Award #DE- FG02-17ER16362"