💊 cMolGPT: A Conditional Generative Pre-Trained Transformer for Target-Specific De Novo Molecular Generation

📝 Medium post

Implementation of "cMolGPT: A Conditional Generative Pre-Trained Transformer for Target-Specific De Novo Molecular Generation". Enforcing target embeddings as queries and keys.

Please feel free to open an issue or email wenlu.wang.1@gmail.com and ye.wang@biogen.com if you have any questions. We will respond as soon as we can.

Dependencies

environment_v100.yml tested on NVIDIA V100

environment_a6000.yml tested on RTX A6000

Create env from yml file

Data

Mol_target_dataloader

Please download this repo and put the folder in the root directory. If you would like to finetune with your own target data, please replace 'target.smi'.

How to run

*unzip train.sim.zip

Train

  python3 main.py --batch_size 512 --mode train \
                  --path model_base.h5 

Fine-tune

  python3 main.py --batch_size 512 --mode finetune \
                  --path model_base.h5 --loadmodel

*In the case of fine-tuning, the base model will be overwritten in place.

*You can change the number of targets in model_auto.py.

Infer/Generate

  python3 main.py --mode infer --target [0/1/2/3] --path model_finetune.h5

No target

  python3 main.py --mode infer --target 0 --path model_finetune.h5

Target 2

  python3 main.py --mode infer --target 2 --path model_finetune.h5