Refining Openfold predictions with Crystallographic/Cryo-EM LiKElihood Targets (ROCKET)

This is the code repo for AlphaFold as a Prior: Experimental Structure Determination Conditioned on a Pretrained Neural Network

You can find detailed documentation and walk-through tutorials at: https://rocket-9.gitbook.io/rocket-docs

Installation

1. Install OpenFold

To ensure usability, we forked the OpenFold repo, and sorted a couple details in the installation guides. Here is what we advise ROCKET users to do:

Note: The ⁠openfold installation requires approximately 6 GB of free space to download weights. Please ensure you start in a directory with sufficient available space.

Note: To ensure a smooth installation and execution of ROCKET, install on a GPU machine that matches the hardware you’ll use in production. In other words, for HPC users, if you plan to run your code on a node with a particular GPU model, request the same GPU model when you install OpenFold. This is important because the installation process performs a hardware-specific compilation. We also recommend using GPUs with CUDA Compute Capability 8.0 or higher.

1. Clone our fork of the OpenFold repo, switch to the pl_upgrades branch to work with CUDA 12:

   git clone https://github.com/minhuanli/rocket_openfold.git
   cd rocket_openfold
   git checkout pl_upgrades
   

2. Create a conda/mamba env with the environment.yml

   Note: If you work with an HPC cluster with package management like module, purge all your modules before this step to avoid conflicts.

   mamba env create -n <env_name_you_like> -f environment.yml
   mamba activate <env_name_you_like>
   

   The main change we made is moving the flash-attn package outside of the yml file, so you can install it manually afterwards. This is necessary because this OpenFold version relies on pytorch 2.1, which is incompatible with the latest flash-attn, so a simple pip install flash-attn would fail. Also using a --no-build-isolation flag allows using ninja for compilation, which is much faster.

3. Install compatible flash-attn (latest flash-attn with noted support for pytroch-2.1 + cuda-12.1)

   pip install flash-attn==2.2.2 --no-build-isolation
   

4. Run the setup script to install OpenFold, and configure kernels and folding resources

   ./scripts/install_third_party_dependencies.sh
   

   Add the following lines to <path_to_your_conda_env>/etc/conda/activate.d/env_vars.sh, create it if it doesn't exist

   #!/bin/sh
   
   export LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH
   export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
   

   This is so everytime you activate this env, the prepend will happen automatically.

5. Download AlphaFold2 weights, add the resources path to system environment (we need this for ROCKET)

   ./scripts/download_alphafold_params.sh ./openfold/resources
   

   Note: You can download OpenFold weights if you want to try

   Append the following line to <path_to_your_conda_env>/etc/conda/activate.d/env_vars.sh, you should have created it from the previous step

   export OPENFOLD_RESOURCES="<ABSOLUTE_PATH_TO_OPENFOLD_FOLDER>/openfold/resources"
   

   <ABSOLUTE_PATH_TO_OPENFOLD_FOLDER> should be the output of pwd -P you get from the OpenFold repo path.

   Deactivate and reactivate your python environment, you should be able to run and see the path:

   echo $OPENFOLD_RESOURCES 
   

6. Check your OpenFold build with unit tests:

   ./scripts/run_unit_tests.sh
   

   Ensure you see no errors:

   ...
   Time to load evoformer_attn op: 243.8257336616516 seconds
   ............s...s.sss.ss.....sssssssss.sss....ssssss..s.s.s.ss.s......s.s..ss...ss.s.s....s........
   ----------------------------------------------------------------------
   Ran 117 tests in 275.889s
   
   OK (skipped=41)
   

2. Install Phenix (required for automatic preprocessing and post-refinement)

Phenix is required for automatic data preprocessing and for post-refinement when polishing final model geometry. Follow the steps below to install it and add the path to the system environment variables:

1. Download the latest nightly-build Phenix python3 installer according to https://phenix-online.org/download, note you have download the installer from the show-all link, with version newer than 2.0rc1-5647

2. Run the installer

   bash phenix-installer-2.0rc1-5617-<platform>.sh
   

   You will be prompted to type your preferred path of installation, after specifying it, you will see:

   Phenix will now be installed into this location:
   <phenix_directory>/phenix-2.0rc1-5617
   

   Note: <phenix_directory> must be a absolute path. The installer will will make <phenix_directory>/phenix-2.0rc1-5617 and install there.

3. Append the following line to <path_to_your_conda_env>/etc/conda/activate.d/env_vars.sh, you should have created it from the previous section

   export PHENIX_ROOT="<phenix_directory>/phenix-2.0rc1-5617"
   

   <phenix_directory> is where you install Phenix in the last step

   Deactivate and reactivate your python environment, you should be able to run and see the path:

   echo $PHENIX_ROOT 
   

3. Install ROCKET

ROCKET is available through pypi as rs-rocket, you can easily install by run

pip install rs-rocket

For developer mode, you can fetch the repo and do editable installation locally

git clone https://github.com/alisiafadini/ROCKET.git
cd ROCKET
pip install -e ".[tests,CI]"

Run rk.score --help after installation, if you see a normal doc strings without errors, you are good to go!

Citing

@article{fadini2025alphafold,
  title={AlphaFold as a Prior: Experimental Structure Determination Conditioned on a Pretrained Neural Network},
  author={Fadini, Alisia and Li, Minhuan and McCoy, Airlie J and Terwilliger, Thomas C and Read, Randy J and Hekstra, Doeke and AlQuraishi, Mohammed},
  journal={bioRxiv},
  year={2025}
}