Title: Exploring the fentanyl chemical space by using deep learning models

The code in this file converts the smiles format of molecules into SeqGAN training data.

The environment configured for this code is in the tf1_py2.yml. In this file is the code for SeqGAN.

conda activate tf1_py2 cd fentanyl analogues generation/code/2、seqGAN/SeqGAN-1 python sequence_gan.py

The code in this file converts SeqGAN generation data into the smiles format of molecules.

The code in this file can pick out the valid molecules from the generated molecules.

The code in this file converts the smiles format for generating molecules into molecular diagrams.

The code in this file can calculate molecular similarity and draw similarity heatmaps.

The code in this file can calculate the properties of molecules and draw the distibution of properties of molecules.

The code in this file can do data augmentation.

The code in this file can remove duplicate molecules.

In this work, we train small custom GPT on Moses and Guacamol dataset with next token prediction task. The model is then used for conditional molecular generation.

• The processed Guacamol and MOSES datasets in csv format can be downloaded from this link:

https://drive.google.com/drive/folders/1LrtGru7Srj_62WMR4Zcfs7xJ3GZr9N4E?usp=sharing

• Original Guacamol dataset can be found here:

https://github.com/BenevolentAI/guacamol

• Original Moses dataset can be found here:

https://github.com/molecularsets/moses

• All trained weights can be found here:

https://www.kaggle.com/virajbagal/ligflow-final-weights

To train the model, make sure you have the datasets' csv file in the same directory as the code files.

./train_moses.sh

./train_guacamol.sh

./generate_guacamol_prop.sh

./generate_moses_prop_scaf.sh

If you find this work useful, please cite:

Bagal, Viraj; Aggarwal, Rishal; Vinod, P. K.; Priyakumar, U. Deva (2021): MolGPT: Molecular Generation using a Transformer-Decoder Model. ChemRxiv. Preprint. https://doi.org/10.26434/chemrxiv.14561901.v1