FREED++

This repository is the official Pytorch implementation of "FREED++: Improving RL Agents for Fragment-Based Molecule Generation by Thorough Reproduction".

Installation

Dependency

The codes have been tested in the following environment:

Package	Version
Python	3.7.12
PyTorch	1.12.1
TorchVision	0.13.1
CUDA	11.3.1
DGL	0.9.1.post1
RDKit	2020.09.1.0

Install via conda yaml file

conda env create -f environment.yml
conda activate ffreed

Install via Dockerfile

docker build -t ffreed .
docker run -it --name ffreed -v /home/user/freed:/home/jovyan --gpus device=0 ffreed

Training

FREED++

python main.py \
    --exp_root /home/user/freed/experiments \
    --alert_collections /home/user/freed/data/alert_collections.csv \
    --fragments /home/user/freed/data/motifs/zinc_crem.json \
    --receptor /home/user/freed/data/receptors/protein.pdbqt \
    --vina_program /home/user/freed/utils/qvina02 \
    --starting_smile "c1([*:1])c([*:2])ccc([*:3])c1" \
    --fragmentation crem \
    --num_sub_proc 12 \
    --n_conf 1 \
    --exhaustiveness 1 \
    --save_freq 50 \
    --epochs 200 \
    --commands "train,sample" \
    --reward_version soft \
    --box_center "x1,x2,x3" \
    --box_size "s1,s2,s3" \
    --seed 150 \
    --name freedpp

FFREED

python main.py \
    --exp_root /home/user/freed/experiments \
    --alert_collections /home/user/freed/data/alert_collections.csv \
    --fragments /home/user/freed/data/motifs/zinc_crem.json \
    --receptor /home/user/freed/data/receptors/protein.pdbqt \
    --vina_program /home/user/freed/utils/qvina02 \
    --starting_smile "c1([*:1])c([*:2])ccc([*:3])c1" \
    --fragmentation crem \
    --num_sub_proc 12 \
    --n_conf 1 \
    --exhaustiveness 1 \
    --save_freq 50 \
    --epochs 200 \
    --commands "train,sample" \
    --reward_version soft \
    --box_center "x1,x2,x3" \
    --box_size "s1,s2,s3" \
    --seed 150 \
    --name ffreed \
    --action_mechanism sfps \
    --per True \
    --merger mi

We recommend to specify timeout_dock and timeout_gen3d parameters, since unconstrained usage of OpenBabel and Qvina2 may slowdown training.

Evaluation

python main.py \
    --exp_root /home/user/freed/experiments \
    --alert_collections /home/user/freed/data/alert_collections.csv \
    --fragments /home/user/freed/data/motifs/zinc_crem.json \
    --receptor /home/user/freed/data/receptors/protein.pdbqt \
    --vina_program /home/user/freed/utils/qvina02 \
    --starting_smile "c1([*:1])c([*:2])ccc([*:3])c1" \
    --fragmentation crem \
    --num_sub_proc 12 \
    --n_conf 3 \
    --exhaustiveness 8 \
    --save_freq 50 \
    --epochs 200 \
    --commands "evaluate" \
    --reward_version soft \
    --box_center "x1,x2,x3" \
    --box_size "s1,s2,s3" \
    --seed 150 \
    --timeout_dock 90 \
    --timeout_gen3d 30 \
    --name ffreed

Citation

To cite this work, please use:

@article{
telepov2023freed,
title={{FREED}++: Improving {RL} Agents for Fragment-Based Molecule Generation by Thorough Reproduction},
author={Alexander Telepov and Artem Tsypin and Kuzma Khrabrov and Sergey Yakukhnov and Pavel Strashnov and Petr Zhilyaev and Egor Rumiantsev and Daniel Ezhov and Manvel Avetisian and Olga Popova and Artur Kadurin},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=YVPb6tyRJu},
note={}
}