Open Macromolecular Genome

This repository contains python scripts to construct the Open Macromolecular Genome (OMG) database from eMolecules and train generative models described in the paper. The OMG monomers and polymers are available at Zenodo.

Setup

Set up Python environment with Anaconda

conda env create -f environment.yml

eMolecules database

To begin, you will need the version.smi file from eMolecules. Download here: https://marketing.emolecules.com/incremental-file-download

Place version.smi in the data directory.

|-data
    |-version.smi

Script components

To run a script, a file path in the script should be modified to be consistent with an attempted directory.

1. data

This directory contains scripts to construct the OMG database from eMolecules.

2. scscore

This directory contains a script to calculate SC score obtained from https://github.com/connorcoley/scscore.

3. polymerization

This directory contains the OMG polymerization algorithms.

4. selfies

This directory contains modified SELFIES scripts to incorporate asterisk (*). The asterisk rules were added to the original work, https://github.com/aspuru-guzik-group/selfies

5. molecule_chef

This directory contains the Molecule Chef generative model. The scripts were written referring to the original work, Bradshaw, J.; Paige, B.; Kusner, M. J.; Segler, M.; Hernández-Lobato, J. M. A Model to Search for Synthesizable Molecules. In Advances in Neural Information Processing Systems; Curran Associates, Inc., 2019; Vol. 32., and their scripts, https://github.com/john-bradshaw/molecule-chef.

6. vae

This directory contains a variational autoencoder model. The scripts were written referring to https://github.com/aspuru-guzik-group/selfies/blob/master/examples/vae_example/chemistry_vae.py

7. train

This directory contains scripts to train Molecule Chef and SELFIES VAE. These scripts were written by referring to https://github.com/john-bradshaw/molecule-chef and https://github.com/aspuru-guzik-group/selfies/blob/master/examples/vae_example/chemistry_vae.py.

Authors

Seonghwan Kim, Charles M. Schroeder, and Nicholas E. Jackson

Funding Acknowledgements

This work was supported by the IBM-Illinois Discovery Accelerator Institute. N.E.J. thanks the 3M Nontenured Faculty Award for support of this research. We thank Jed Pitera and Jeffrey Moore for critical readings of the manuscript and Prof. Tengfei Luo for assistance with the PI1M dataset.