AFEXplorer-min: The minimum implementation of AFEXplorer for conditioned prediction of protein structures.

Zilin Song (song.zilin@outlook.com)

Installation

If you have a working AF2 environment, install the following packages depends on your JAX/JAXLIB version in the environment. For more info, see the corresponding GitHub repos and find a compatible release.

chex
optax

If installing from scratch, the following provides a non-Docker approach, depends on your cuda version (below are for CUDA 11.2):

conda create --name af2
conda activate af2
conda install -y -c conda-forge openmm==7.5.1 cudatoolkit=11.2 pdbfixer
pip install -r requirements.txt --no-cache-dir
pip install --upgrade --no-cache-dir jax==0.3.25 jaxlib==0.3.25+cuda11.cudnn805 optax==0.1.7 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

If an error related to "ptxas version string" arises, try module load your cuda library.

NOTE - Z.S.

• This implementation is compatible to both AF and AF-Multimer.
• I have only made the following changes to the AF source code:
  alphafold/model/modules_multimer.py lines 427-429, 520-522, 647-649.
• In principle, the AFEX code can also be migrated to ColabFold (which essentially enables more efficient MSA featurizations and shares the identical inference pipeline with AF/AF-Multimer). However, we did not test AFEX with ColabFold at this moment.

References

If referring to the AF-Monomer implementation of AFEX, you are welcomed to cite:

@article{song2024afex,
  title={Conditioned protein structure prediction},
  author={Xie, Tengyu and Song, Zilin and Huang, Jing},
  journal={PRX Life},
  volume={2},
  number={4},
  pages={043001},
  year={2024},
  publisher={APS}
}