This is the official implementation of junction tree encoder-decoder model in https://arxiv.org/abs/1812.01070

• Python == 2.7
• RDKit >= 2017.09
• PyTorch >= 0.4.0
• Numpy
• scikit-learn

The code has been tested under python 2.7 with pytorch 0.4.1.

The tutorial of training and testing our variational junction tree encoder-decoder is in diff_vae/README.md.

A quick summary of different folders:

• data/ contains the training, validation and test set of logP, QED and DRD2 tasks described in the paper.
• fast_jtnn/ contains the implementation of junction tree encoder-decoder.
• diff_vae/ includes the training and decoding script of variational junction tree encoder-decoder (README).
• diff_vae_gan/ includes the training and decoding script of adversarial training module (README).
• props/ is the property evaluation module, including penalized logP, QED and DRD2 property calculation.
• scripts/ provides evaluation and data preprocessing scripts.

Wengong Jin ([email protected])