Metro: Memory-Enhanced Transformer for Retrosynthetic Planning via Reaction Tree

This repository contains an implementation of "Metro: Memory-Enhanced Transformer for Retrosynthetic Planning via Reaction Tree".

Note that we provide the model files for all models, you can skip the training and use our provided files for testing directly.

Dropbox

We provide the starting material file in dropbox, you can download this file via: https://www.dropbox.com/s/nwh2ijrjzbyia73/zinc_stock_17_04_20.hdf5?dl=0
Please move this file into the root folder.

Metro

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 Metro/  
cd Metro  

#Data Process  
python to_canolize.py  

#Train 
python train.py --batch_size 64 --epochs 4000  

#Test 
python top5_inference.py  

Transformer

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 Transformer/  
cd Transformer  

#Data Process  
python to_canolize.py  

#Train  
python train.py --batch_size 64 --epochs 2000  

#Test  
python top5_inference.py  

Retrosim

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 Retrosim/  
cd Retrosim  

#Test
python top5_inference.py  

Neuralsym

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 Neuralsym/  

#Model download
Download the model folder via:  
https://www.dropbox.com/sh/lw8e6epk74vmsh0/AADr0VhLiA6LwC34_Qj0Naufa?dl=0  
mv checkpoint Neuralsym/  
cd Neuralsym  

#Data Process
python prepare_data.py  

#Train
bash train.sh  

#Test
python top5_inference.py  

GLN

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 GLN/gln/  
cd GLN  
pip install -e  
cd gln  

#Data Process 
python process_data.py -save_dir data -num_cores 12 -num_parts 1 -fp_degree 2 -f_atoms data/atom_list.txt -retro_during_train False $@  

#Train
bash run_mf.sh schneider  

#Test
python top5_inference.py -save_dir data -f_atoms data/atom_list.txt -gpu 0  

Megan

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 Megan/data/  

#Model download
Download the model folder via:  
https://www.dropbox.com/sh/aq9h5jyautjpgrh/AAAv_cC2CiQtYTbP7IKFqQcMa?dl=0  
mv models Megan/  
cd Megan  
source env.sh  

#Data Process  
python json2csv.py  
python acquire.py uspto_50k  
python featurize.py uspto_50k megan_16_bfs_randat  

#Train
python bin/train.py uspto_50k models/uspto_50k  

#Test
python bin/top5_inference.py models/uspto_50k --beam-size 5  

GraphRetro

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 GraphRetro/datasets/uspto-50k  
cd GraphRetro  
export SEQ_GRAPH_RETRO=/path/to/dir/  
python setup.py develop  

#Data Process
mv datasets/uspto-50k/valid_data.json datasets/uspto-50k/eval_data.json  
python json2csv.py  
python data_process/canonicalize_prod.py --filename train.csv  
python data_process/canonicalize_prod.py --filename eval.csv  
python data_process/canonicalize_prod.py --filename test.csv  
python data_process/core_edits/bond_edits.py  
python data_process/lg_edits/lg_classifier.py  
python data_process/lg_edits/lg_tensors.py  

#Train
python scripts/benchmarks/run_model.py --config_file configs/single_edit/defaults.yaml  
python scripts/benchmarks/run_model.py --config_file configs/lg_ind/defaults.yaml  

#Test
python scripts/eval/top5_inference.py --edits_exp SingleEdit_20220823_044246 --lg_exp LGIndEmbed_20220823_04432 --edits_step best_model --lg_step best_model --exp_dir models  

G2Gs

cp train_dataset.json valid_data.json test_data.json zinc_stock_17_04_20.hdf5 G2Gs/datasets/  

#Model download
Download the model files via:  
https://www.dropbox.com/s/bvrsun8yap4sal9/g2gs_reaction_model.pth?dl=0  
https://www.dropbox.com/s/0n2ff4zbnbml6xq/g2gs_synthon_model.pth?dl=0  
mv g2gs_synthon_model.pth g2gs_reaction_model.pth G2Gs/  
cd G2Gs  

#Train
python script/train.py -g [0]  

#Test
python script/top5_inference.py -g [0] -k 5 -b 1  

Reference

Retrosim: https://github.com/connorcoley/retrosim
Neuralsym: https://github.com/linminhtoo/neuralsym
GLN: https://github.com/Hanjun-Dai/GLN
G2Gs: https://torchdrug.ai/docs/tutorials/retrosynthesis
GraphRetro: https://github.com/vsomnath/graphretro
Transformer: https://github.com/bigchem/synthesis
Megan: https://github.com/molecule-one/megan

Citation

@article{liu2022metro,
  title={Metro: Memory-Enhanced Transformer for Retrosynthetic Planning via Reaction Tree},
  author={Liu, Songtao and Ying, Rex and Zhang, Zuobai and Zhao, Peilin and Lin, Lu and Wu, Dinghao},
  journal={arXiv preprint arXiv:2209.15315},
  year={2022}
}