Skip to content

Inference code for PoET: A generative model of protein families as sequences-of-sequences

License

Notifications You must be signed in to change notification settings

OpenProteinAI/PoET

Repository files navigation

PoET: A generative model of protein families as sequences-of-sequences

This repo contains inference code for "PoET: A generative model of protein families as sequences-of-sequences", a state-of-the-art protein language model for variant effect prediction and conditional sequence generation.

Environment Setup

  1. Have mamba (faster alternative to conda) installed (Instructions)
  2. Have conda-lock installed in your base conda/mamba environment (Instructions)
  3. Run make create_conda_env. This will create a conda environment named poet.
  4. Run make download_model to download the model (~400MB). The model will be located at data/poet.ckpt. Please note the license.

Scoring variants

Use the script scripts/score.py to obtain fitness scores for a list of protein variants given a MSA of homologs of the WT sequence.

  1. Be on a machine with a NVIDIA GPU. The model cannot run on CPU only.

  2. Activate the poet conda environment

  3. Run the script, replacing the values in angle brackets with the appropriate paths.

    python scripts/score.py \
    --msa_a3m_path <path to MSA of homologs of WT sequence> \
    --variants_fasta_path <path to fasta file containing variants to score> \
    --output_npy_path <path to output file where scores for each variant will be stored as a numpy array>
    

You can pass a lower value for the batch size (--batch_size) if you run out of VRAM. The script was tested on an A100 GPU with 40GB VRAM.

Example

Run the scoring script without arguments python scripts/score.py to score variants in the BLAT_ECOLX_Jacquier_2013 dataset from ProteinGym.

  • the dataset is located at data/BLAT_ECOLX_Jacquier_2013.csv
  • the variants to score as a fasta file is located at data/BLAT_ECOLX_Jacquier_2013_variants.fasta
  • the MSA of homologs of the WT sequence, generated using ColabFold MMseqs2 with the UniRef2202 database, is located at data/BLAT_ECOLX_ColabFold_2202.a3m
  • the scores will be saved as a numpy array at data/BLAT_ECOLX_Jacquier_2013_variants.npy

The scores obtained from the script should obtain >0.65 Spearman correlation with the measured fitness (DMS_score column in the dataset file).

Citation

You may cite the paper as

@inproceedings{NEURIPS2023_f4366126,
 author = {Truong Jr, Timothy and Bepler, Tristan},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {A. Oh and T. Neumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},
 pages = {77379--77415},
 publisher = {Curran Associates, Inc.},
 title = {PoET: A generative model of protein families as sequences-of-sequences},
 url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/f4366126eba252699b280e8f93c0ab2f-Paper-Conference.pdf},
 volume = {36},
 year = {2023}
}

License

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

The PoET model weights (DOI: 10.5281/zenodo.10061322) are available under the CC BY-NC-SA 4.0 license for academic use only. The license can also be found in the LICENSE file provided with the model weights. For commercial use, please reach out to us at [email protected] about licensing. Copyright (c) NE47 Bio, Inc. All Rights Reserved.

Releases

No releases published

Packages

No packages published