Awesome

Paper Info.

Hyunseung Kim, Jonggeol Na*, and Won Bo Lee*, "Generative Chemical Transformer: Neural Machine Learning of Molecular Geometric Structures from Chemical Language via Attention, " J. Chem. Inf. Model. 2021, 61, 12, 5804-5814
[Featured as JCIM journal supplentary cover] https://pubs.acs.org/doi/abs/10.1021/acs.jcim.1c01289

Generative Chemical Transformer (GCT) directly designs hit-like materials matched given target properties. Transformer's high level of context-awareness ability of chemical language aids in designing molecules more realistic and vaild.

Citation

@article{kim2021generative,
  title={Generative Chemical Transformer: Neural Machine Learning of Molecular Geometric Structures from Chemical Language via Attention},
  author={Kim, Hyunseung and Na, Jonggeol and Lee, Won Bo},
  journal={Journal of Chemical Information and Modeling},
  volume={61},
  number={12},
  pages={5804--5814},
  year={2021},
  publisher={ACS Publications}
}

molGCT: Transformer+cVAE

Overview of GCT: Tranining phase and Inference phase
<img src = "images/image1.svg" width="75%">

Design hit-like materials with 10-sets (#1~#10) of given target properties
<img src = "images/image6.jpg" width="75%">

Attention mechanism aids in undertanding semantically related parts of molecular structure hidden in 1-dimensional chemical language
<img src = "images/image3.svg" width="75%">

GCT generates more realistic and valid based on the high level of chemical language comprehension of Transformer:
The molecules in circle are real molecules in MOSES database, and the molecules outside the circle are designed by GCT

Result examples
<img src = "images/image2.png" width="85%">

Implementation

This code was tested in Windows OS

Set up your anaconda environment with the following code:

conda env create -f molgct_env.yaml

Run to train GCT:

python train.py

Run to infer molecules with the trained GCT:

python inference.py

References

Basic Transformer code was borrowed and modified from: https://github.com/SamLynnEvans/Transformer
Molecular data were borrowed from: https://github.com/molecularsets/moses/tree/master/data