The official repository of our paper "Deep reinforcement learning as an interaction agent to steer fragment-based 3D molecular generation for protein pockets". The complete repository is under construction (updated on April 17, 2024).
- We have presented the conda environment file in
./environment.yml
. - We have evaluated our models use external tools, including: Qvina, Pyscreener.
- We pre-trained our model using the natural product dataset COCONUT and the Pocket3D dataset collected from the Protein Data Bank. The dataset used for fine-tuning was obtained from CrossDocked2020.
- To facilitate your implementation, we have provided the raw datasets used by AMG. Download the dataset archive from AMG-DATA.
Ligand encoder and fragment-based decoder pre-training:
python pretrain_ligand.py
Pocket encoder pre-training:
python pretrain_pocket.py
The first training stage:
python train_rec.py
The second training stage:
python train_agent.py
python sample.py