AFEXplorer

AFEXplorer (AFEX) is a generic approach that tailors AlphaFold predictions to user-defined constraints in coarse coordinate spaces by optimizing embedding features. It has been demonstrated to be effective in predicting protein alternative conformations conditioned by local or global structure features for human kinases and membrane transporters.

Lead developer: Zilin Song (song.zilin@outlook.com)

Installation

Assume you have a working AlphaFold 2.3 environment, and install the following packages depending on your JAX/JAXLIB version in the environment. For more info, see the corresponding GitHub repos and find a compatible release.

chex
optax

Usage of AFEX: AdK protein as an example

1. Prepare input.

   Run AlphaFold for features.pkl, which is the input of AFEX.

2. Conditioned structure prediction.

   • Add corresponding CV loss for the open state in afexplore_optim.py.

   • Set the directory of AF model parameters as $AF_PARAM.

   • Run AFEX.

     cd scripts
     bash run_afexplore_optim.sh ../data_afexplore_monomer_ADKopen ADK out $AF_PARAM 

3. The output put will be in data_afexplore_monomer_ADKopen/ADK.

4. It is flexible for user to adjust the AFEX optimizaiton-related parameters (e.g. learning rate) in afexplore_optim.py and run_afexplore_optim.sh.

Additional Notes

Kincore-standalone is also required when predicting human kinases. Simply, hmmer needs to be installed in the current environment to support Kincore. See Kincore-standalone for more info.

Currently, AFEX is compatible with AF-Monomer.

Reference

Tengyu Xie#, Zilin Song#, Jing Huang. Conditioned Protein Structure Prediction. bioRxiv 2023.09.24.559171; doi: https://doi.org/10.1101/2023.09.24.559171.

Acknowledgements

Inspiration, code snippets, etc.

• AlphaFold
• AFProfile