This repo contains the scripts described in the paper Improving de novo Protein Binder Design with Deep Learning.
This repository provides a copy of Brian Coventry's silent tools, these tools are also provided here. This repository provides a wrapper to Justas' ProteinMPNN code and some of the code in the wrapper class is adapted directly from ProteinMPNN code. Finally, this repository provides a version of the AlphaFold2 source code with the "initial guess" modifications described in this paper. The AF2 source code is provided with the original DeepMind license at the top of each file.
I have split the single conda env that this repo used to use (dl_binder_design.yml; still provided in <base_dir>/include if anyone is interested) into two smaller and easier to install environments. The old environment required PyTorch, JAX, TensorFlow, and PyRosetta packages to all be compatible with one another, which is difficult. The new environments should be easier to install and I have also added import tests so that it is easier and faster to check that the installation has been successful.
The steps to installing the environments are as follows:
- Ensure that you have the Anaconda or Miniconda package manager
- Clone this repo
- Navigate to <base_dir>/include
- Run
conda env create -f proteinmpnn_fastrelax.yml - Test the environment by activating your environment and running
python importtests/proteinmpnn_importtest.py, if you encounter an error in this script then something has gone wrong with your installation. If the script prints out that the tests pass then you have installed this environment correctly. This script will also test whether it can access a GPU, it is not recommended to run ProteinMPNN with a GPU as it is only marginally faster than running on CPU and FastRelax cannot take advantage of the GPU anyway.
- Navigate to <base_dir>/include
- Run
conda env create -f af2_binder_design.yml - Test the environment by activating your environment and running
python importtests/af2_importtest.py. Run this script from a node which has acccess to the GPU you wish to use for AF2 inference, this script will print out a message about whether it was able to find and use the GPU on your node. If this script hits an error before printing anything then the installation has not been done correctly.
Getting a conda environment that recognizes and can run AF2 on your GPU is one of the more difficult parts of this process. Because of the many different GPUs out there, it is not possible for us to provide one .yml file that will work with all GPUs. We provide a CUDA 11.1 compatible env (dl_binder_design.yml) and a CUDA 12 compatible env (af2_binder_design.yml). For other versions of the CUDA driver, you may need to change which CUDA version is installed in the conda env, this can be done by changing the CUDA version in this line of af2_binder_design.yml: - jax[cuda12_pip].
NOTE: This conda environment can only accomodate NVIDIA GPUs at this point in time.
This repo requires the code from ProteinMPNN to work. It expects this code to be in the mpnn_fr directory so we can just clone it to be there
- Naviate to <base_dir>/mpnn_fr
- Run
git clone https://github.com/dauparas/ProteinMPNN.git
The scripts contained in this repository work with a type of file called silent files. These are essentially a bunch of compressed .pdb files that are all stored in one file. Working with silent files is conventient and saves a lot of disk space when dealing with many thousands of structures.
Brian Coventry wrote a bunch of really nice commandline tools (called silent_tools) to manipulate silent files. These tools are included in this repository but may also be downloaded separately from this GitHub repo.
The two commands that allow you to go from pdb to silent file and back are the following:
pdbs to silent: <base_dir>/silentfrompdbs *.pdb > my_designs.silent
silent to pdbs: <base_dir>/silentextract all_structs.silent
NOTE: Some silent tools require PyRosetta and will fail if run in a Python environment without access to PyRosetta.
The scripts in this repository expect AF2 weights to be in <base_dir>/model_weights/params and will fail to run if the weights are not there. If you already have AF2 params_model_1_ptm.npz weights downloaded then you may simply copy them to <base_dir>/model_weights/params or create a symlink. If you do not have these weights downloaded you will have to download them from DeepMind's repository, this can be done as follows:
cd <base_dir>/af2_initial_guess
mkdir -p model_weights/params && cd model_weights/params
wget https://storage.googleapis.com/alphafold/alphafold_params_2022-12-06.tar
tar --extract --verbose --file=alphafold_params_2022-12-06.tar
The binder design pipeline requires protein binder backbones as an input. The recommended way to generate these backbones is to use RFdiffusion, which will give you a directory of .pdb files. These files can either be turned into silent files which are more memory efficient and easier to work with than a directory of pdb files (and your system administrator will thank you for using them), or you can use these directories of .pdb files as-is with this pipeline, using either the -pdbdir flag alone or the -pdbdir flag in combination with the -runlist flag.
Example commands demonstrating how to run each of these scripts with different types of input can be found here:
<base_dir>/examples
Here is an example of how to run ProteinMPNN-FastRelax with a silent file of designs. This will use the default of 1 FastRelax cycle of 1 ProteinMPNN sequence per round (NOTE: running with -relax_cycles > 0 and -seqs_per_struct > 1 is disabled as it leads to an explosion in the amount of FastRelax trajectories being run and is probably bad idea):
<base_dir>/mpnn_fr/dl_interface_design.py -silent my_designs.silent
This will create a file titled out.silent containing your designs. This file can be fed directly to AF2 interface prediction.
With the refactor, this script is now able to read and write both PDB files and silent files. The script I have also added more informative argument messages which can be accessed by running:
<base_dir>/mpnn_fr/dl_interface_design.py -h
NOTE: This script expects your binder design to be the first chain it receives. This script is robust to non-unique indices, unlike the AF2 interface script. NOTE 2: The outputs of this script do not have relaxed sidechains (since sidechains are not input to AF2 and it's not worth the computation to relax them) so the structures will look strange if you visualize them in PyMol, this is perfectly normal, the structures will look better after run though AF2.
If you used RFdiffusion to generate your binder designs and would like to fix a region, you can use the following command to add 'FIXED' labels to your pdbs which will be recognized by the ProteinMPNN scripts. Thanks to Preetham Venkatesh for writing this!
python <base_dir>/helper_scripts/addFIXEDlabels.py --pdbdir /dir/of/pdbs --trbdir /dir/of/trbs --verbose
These pdb files can be collected into a silent file (or just used as PDB files) and run through the ProteinMPNN script which will detect the FIXED labels and keep those sequence positions fixed.
Running the interface prediction script is simple:
<base_dir>/af2_initial_guess/predict.py -silent my_designs.silent
This will create a file titled out.silent containing the AF2 predictions of your designs. It will also output a file titled out.sc with the scores of the designs, pae_interaction is the score that showed the most predictivity in the experiments performed in the paper.
With the refactor, this script is now able to read and write PDB and silent files as well as perform both monomer and complex predictions. The arguments for these can be listed by running:
<base_dir>/af2_initial_guess/predict.py -h
NOTE: This script expects your binder design to be the first chain it receives. The binder will be predicted from single sequence and with an intial guess. The target chains will be fixed to the input structure. The script also expects your residue indices to be unique, ie. your binder and target cannot both start with residue 1.
One of the most common errors that people have been having is one that looks like this:
Struct with tag SAMETAG failed in 0 seconds with error: <class 'EXCEPTION'>
Where SAMETAG and EXCEPTION can be many different things. What is happening here is that the main loops of both of the scripts provided here are wrapped in a try-catch block; the script tries to run each design and if an error occurs, the script notes which design had an error and continues to the next design. This error catching is convenient when running production-scale design campaigns but is a nuisance for debugging since the messages are not very informative.
If you hit this error, I recommend running the same command that yielded the error but while adding the -debug flag to the command. This flag will make the script run without the try-catch block and errors will print with the standard verbose, easier-to-debug messages.