Retro-R1 for Retrosynthetic Planning

This repository is based on the open-source codebase of Retro* and Agent-R1.

Setup

Create conda environment

cd Retro-R1
conda env create --file environment.yml
conda activate Retro_R1
pip install flash-attn==2.7.4.post1
pip install -e packages/mlp_retrosyn
pip install -e packages/rdchiral
unzip verl.zip
cd verl
pip install --no-deps -e .

Download the necessary files

To reproduce the results in the paper, we also need the additional files containing the training dataset, evaluation datasets (USPTO, ChEMBL-1000), starting molecules, and the template rules. Files from Retro* (such as single-step model weights V1) can be downloaded from the link, and put the folders (dataset/, one_step_model/) under this directory. The single-step model weights V2 and V3 can be downloaded from the link and put the files (retro_star_value_ours.ckpt(V2) and retro_star_zero_ours.ckpt(V3)) under the one_step_model/ directory. The single-step model weights V4 is not released and can be trained following the instructions in PDVN. We can also provide the trained weights if requested. Put the trained weights under the one_step_model/ directory and name it as retro_star_V4.ckpt. The ChEMBL-1000 testset can be downloaded from link and put under the folder dataset/.

Preprocess the dataset

python ./examples/data_preprocess/reaction.py

Training

bash training_script.py

Note that this script only use one node with eight gpus. To use multi nodes, please follow the instructions of verl.

Export model

Modify the scripts export.sh and change the ori_pth, ckpt_pth and export_pth to export model from the best checkpoint.

bash export.py

Evaluation on Retro*-190 testset

bash test_retro_script.py

Evaluation on ChEMBL-1000 testset

bash test_chembl_script.py