Skip to content

Official repository of PhyloGFN: Phylogenetic inference with generative flow networks

License

Notifications You must be signed in to change notification settings

zmy1116/phylogfn

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PhyloGFN: Phylogenetic inference with generative flow networks

Official repository of PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024)

Mingyang Zhou, Zichao Yan, Elliot Layne, Nikolay Malkin, Dinghuai Zhang, Moksh Jain, Mathieu Blanchette, Yoshua Bengio

diagram

We design a GFlowNet based generative model for phylogenetic inference, achieving strong results in both Bayesian and parsimony-based phylogenetic inference.

Citation

@inproceedings{
    zhou2024phylogfn,
    title={Phylo{GFN}: Phylogenetic inference with generative flow networks},
    author={Ming Yang Zhou and Zichao Yan and Elliot Layne and Nikolay Malkin and Dinghuai Zhang and Moksh Jain and Mathieu Blanchette and Yoshua Bengio},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=hB7SlfEmze}
}

Environment Setup

conda create -n phylogfn python=3.10
conda activate phylogfn
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install anaconda::docopt
conda install etetoolkit::ete3
conda install matplotlib tqdm dill fvcore iopath docopt

Usage

Our latest progress with continuous branch lengths modeling is implemented. To train a phyogfn model

python train.py <cfg_path> <sequences_path> <output_path> [--nb_device=<device_num>] [--quiet] [--amp]
  • Example sequences datasets DS1-DS8 are stored in the folder dataset/benchmark_datasets
  • Example training <cfg_path> are stored in the folder src/configs/benchmark/dna_cfgs
    • Continuous branch lengths modeling configs are in the folder continuous_branch_lengths_modeling
    • Discrete branch lengths modeling configs are in the folder discrete_branch_lengths

PhyloGFN with continuous branch length modeling achieves SOTA MLL estimation performance.

Training trajs DS1 DS2 DS3 DS4 DS5 DS6 DS7 DS8 cfg/weights
PhyloGFN Full 3.20E+07 -7108.4 (0.04) -26367.7 (0.04) -33735.1 (0.02) -13329.9 (0.09) -8214.4 (0.16) -6724.2 (0.10) -37331.9 (0.14) -8650.5 (0.05) -
PhyloGFN 40% 1.28E+07 -7108.4 (0.05) -26367.7 (0.05) -33735.2 (0.04) -13330.1 (0.07) -8214.5 (0.14) -6724.3 (0.10) -37332.1 (0.27) -8650.4 (0.16) -
PhyloGFN 24% 7.68E+06 -7108.4 (0.05) -26367.7 (0.02) -33735.1 (0.07) -13330.0 (0.08) -8214.5 (0.13) -6724.2 (0.21) -37332.2 (0.26) -8650.4 (0.15) googledrive

For discrete branch lengths modeling, we suggest to use the config file with the following configurations:

DS1 DS2 DS3 DS4 DS5 DS6 DS7 DS8
Bin size 0.001 0.004 0.004 0.002 0.002 0.001 0.001 0.001
Bin num 50 50 50 100 100 100 200 100

We will publish the optimized code for parsimony analysis in the near future. In the meantime, if you are interested, please refer to the supplementary materials to run parsimony analysis.

TODO list

  • mutigpu training
  • parsimony inference

About

Official repository of PhyloGFN: Phylogenetic inference with generative flow networks

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages