Installing ROCKET

How to get started

1. Install OpenFold

We forked OpenFold and cleaned up a few installation details for ROCKET users.

Plan for a GPU machine, enough free disk space, and a clean environment before you start.

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 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 or mamba environment with environment.yml. Mamba is recommended because the dependency set is large. If you use Conda instead, replace mamba with conda in the commands below.

   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>

   We keep flash-attn outside the YAML so you can install a compatible version manually. This OpenFold branch uses PyTorch 2.1. The latest flash-attn will not work. The --no-build-isolation flag also lets ninja speed up compilation.

3. Install a compatible flash-attn build:

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

4. Run the setup script:

   ./scripts/install_third_party_dependencies.sh

   Add the following lines to <path_to_your_conda_env>/etc/conda/activate.d/env_vars.sh. Create the file if it does not exist.

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

   This makes the library path update happen automatically whenever the environment is activated.

5. Download AlphaFold2 weights and set the resource path:

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

   Note: You can also download OpenFold weights if you want to test them.

   Append the following line to the same env_vars.sh file:

   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 environment. Then confirm 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.

Use an absolute install path for Phenix. It avoids path issues later.

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 Python 3 nightly build from https://phenix-online.org/download. Use at least version 2.0-5936.

2. Run the installer

   bash phenix-installer-2.0rc1-5936-<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-2.0-5936

   Note: <phenix_directory> must be an absolute path. The installer will create <phenix_directory>/phenix-2.0-5936 and install there.

3. Append the following line to the same env_vars.sh file:

   export PHENIX_ROOT="<phenix_directory>/phenix-2.0-5936"

   <phenix_directory> is the path you chose in the previous step.

   Deactivate and reactivate your environment. Then confirm the path:

   echo $PHENIX_ROOT 

3. Install ROCKET

Move to a separate parent directory, clone ROCKET, then install it with pip:

git clone https://github.com/alisiafadini/ROCKET.git
cd ROCKET
pip install .

This also installs dependencies such as SFcalculator and reciprocalspaceship.

Note: If you get errors about incompatibility of prompt_toolkit, ignore them.

For editable install mode, run:

pip install -e .

Run rk.score --help after installation. If it prints normal help text without errors, the install is ready.

Once this works, move on to the X-ray or cryo-EM setup tutorial.

Create an issue in our repo