Awesome

E(3) Equivariant Translation VAE for 3D Molecular Design

This repo contains the training and generation code for the model described in the paper, "Efficient 3D Molecular Design with an E(3) Invariant Transformer VAE".

Installation

We build the model architecture of Vagrant using PyTorch. Certain functions depend on RDKit and we use Morfeus for conformer generation. We also use standard python packages numpy, pandas, scikit-learn, and scipy. Otherwise the repo is entirely self-contained.

Data

The training script automatically downloads the QM9 dataset and stores it in qm9/temp/qm9 the first time it is run. You can force a redownload by adding the --force_download argument or reprocess the data by adding the --reprocess_data argument. We also include a pretrained checkpoint file in checkpoints/vagrant.

Training

To train a model, run the following code in a terminal python scripts/train.py. The default hyperparameters match the hyperparameters used to train the version of Vagrant reported in the submission.

Generation

You can generate molecules by calling python scripts/gen.py --name vagrant --ckpt_epoch 100. The --name and --ckpt_epoch arguments are rquired. Additional options can be passed to select a sampling method or additional hyperparameters. For instance, to use robust sampling and calculate the coherence of each sample, you would call python scripts/gen.py --name vagrant --ckpt_epoch 100 --sample_method robust --calc_coherence.

Conformers

We also include code for generating structural conformers from the sampled molecules. These functions can be found in vagrant/conformers.py. Generating conformers for some samples can be slow. This may be the case when calculating coherence and can some times cause errors during generation.