Skip to content

Latest commit

 

History

History
34 lines (24 loc) · 1.83 KB

README.md

File metadata and controls

34 lines (24 loc) · 1.83 KB
Raw

Junction Tree Variational Autoencoder for Molecular Graph Generation

Official implementation of our Junction Tree Variational Autoencoder https://arxiv.org/abs/1802.04364

Update

We have made architecture improvements to JT-VAE. We recommend you to check our new repository at https://github.com/wengong-jin/hgraph2graph/. This repo contains a molecular language model pre-trained on ChEMBL (1.8 million compounds) and scripts for property-guided molecule generation. All scripts are written in python 3.7 and pytorch.

Accelerated Version

We have accelerated our code! The new code is in fast_jtnn/, and the VAE training script is in fast_molvae/. Please refer to fast_molvae/README.md for details.

Requirements

  • Linux (We only tested on Ubuntu)
  • RDKit (version >= 2017.09)
  • Python (version == 2.7)
  • PyTorch (version >= 0.2)

To install RDKit, please follow the instructions here http://www.rdkit.org/docs/Install.html

We highly recommend you to use conda for package management.

Quick Start

The following directories contains the most up-to-date implementations of our model:

  • fast_jtnn/ contains codes for model implementation.
  • fast_molvae/ contains codes for VAE training. Please refer to fast_molvae/README.md for details.

The following directories provides scripts for the experiments in our original ICML paper:

  • bo/ includes scripts for Bayesian optimization experiments. Please read bo/README.md for details.
  • molvae/ includes scripts for training our VAE model only. Please read molvae/README.md for training our VAE model.
  • molopt/ includes scripts for jointly training our VAE and property predictors. Please read molopt/README.md for details.
  • jtnn/ contains codes for model formulation.

Contact

Wengong Jin ([email protected])