Skip to content

GraphAF

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

GraphAF

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
config
config
 
 
figs
figs
 
 
README.md
README.md
 
 
run_const_prop_opt.py
run_const_prop_opt.py
 
 
run_prop_opt.py
run_prop_opt.py
 
 
run_rand_gen.py
run_rand_gen.py
 
 

README.md

GraphAF

This is a re-implementation for GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation.

Image

Table of Contents

  1. Usage
  2. Citation
  3. Acknowledgement

Usage

Random Generation

You can use our trained models or train the model from scratch:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_rand_gen.py --train --data=qm9 
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_rand_gen.py --train --data=zinc250k

To generate molecules using trained model, first download models from this link, then:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_rand_gen.py --num_mols=100 --model_path=${path_to_the_model} --data=qm9
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_rand_gen.py --num_mols=100 --model_path=${path_to_the_model} --data=zinc250k

Property Optimization

For property optimization, we aim to generate molecules with desirable properties (i.e., QED and plogp in this work). You can use our trained models or train the model from scratch by reinforcement learning:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_prop_optim.py --train --prop=plogp
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_prop_optim.py --train --prop=qed

To generate molecules using our trained model, first download models from this link, then:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_prop_optim.py --num_mols=100 --model_path=${path_to_the_model} --prop=plogp
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_prop_optim.py --num_mols=100 --model_path=${path_to_the_model} --prop=qed

Constrained Optimization

For constrained optimization, we aim to optimize molecules with desirable properties (plogp in this work). You can use trained models or train the model from scratch by reinforcement learning:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_cons_optim.py --train --data=graphaf

To optimize molecules using trained model, first download models from this link, then:

$ cd examples/ggraph/GraphAF
$ CUDA_VISIBLE_DEVICES=${your_gpu_id} python run_cons_optim.py --model_path=${path_to_the_model} --data=graphaf

Citation

@article{shi2020graphaf,
  title={{GraphAF}: a Flow-based Autoregressive Model for Molecular Graph Generation},
  author={Chence Shi and Minkai Xu and Zhaocheng Zhu and Weinan Zhang and Ming Zhang and Jian Tang},
  journal={iclr},
  year={2020}
}

Acknowledgement

Our implementation is based on GraphAF. Thanks a lot for their awesome works.