A PyTorch implementation of "TenGAN: Pure Transformer Encoders Make an Efficient Discrete GAN for De Novo Molecular Generation." The paper has been accepted by AISTATS 2024.

Execute the following commands:

$ conda env create -n tengan_env -f env.yml
$ source activate tengan_env

• dataset: contains the training datasets. Each dataset contains only one column of SMILES strings.

  • QM9.csv
  • ZINC.csv

• res: all generated datasets, saved models, and experimental results are saved in this folder.

  • save_models: all training results, pre-trained and trained filler and discriminator models are saved in this folder.

  • main.py: definite all hyper-parameters, pretraining of the generator, pretraining of the discriminator, adversarial training of the TenGAN and Ten(W)GAN.

  • mol_metrics.py: definite the vocabulary, tokenization of SMILES strings, and all the objective functions of the chemical properties.

  • data_iter.py: load data for the generator and discriminator.

  • generator.py: definite the generator.

  • discriminator.py: definite the discriminator.

  • rollout.py: definite the Monte Carlo method.

  • utils.py: definite the performance evaluation methods of the generated molecules, such as the validity, uniqueness, novelty, and diversity.

- solubility
- druglikeness
- synthesizability

• TenGAN on the ZINC dataset with drug-likeness as the optimized property:

$ python main.py

C. Li and Y. Yamanishi (2024). TenGAN: Pure transformer encoders make an efficient discrete GAN for de novo molecular generation. AISTATS 2024.

BibTeX format:

@inproceedings{li2024tengan,
title={TenGAN: Pure Transformer Encoders Make an Efficient Discrete GAN for De Novo Molecular Generation},
author={Li, Chen and Yamanishi, Yoshihiro},
booktitle={27th International Conference on Artificial Intelligence and Statistics (AISTATS)},
volume={２３８},
year={2024}
}