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 [email protected] and [email protected] if you have any questions. We will respond as soon as we can.

environment_v100.yml tested on NVIDIA V100

environment_a6000.yml tested on RTX A6000

Create env from yml file

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'.

*unzip train.sim.zip

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

  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.

  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