Fast GPU-based protein folding simulations that run from only a protein sequence and set of restraints designed to fold the protein.

Pre-install
If you do not have a suitable Python 3 installation, we recommend you use an Anaconda Python distribution. If you are installing on a cluster or HPC, you will likely need to specify a non-standard miniconda directory to use the software to avoid permissions issues. For example on OLCF resources such as Summit, a good place to install software is /ccs/proj/<projID>/miniconda3. Here is a lightweight version that will get you started:

# Use this if you are on PPC64LE
wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-ppc64le.sh

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
rm Miniconda3-latest-Linux-x86_64.sh

# If you choose to keep your bashrc clean, add the conda/bin to PATH
#  - change the "pwd" here according to your miniconda install path
export PATH="$(pwd)/miniconda3/bin:$PATH"

# Two dependencies for the installation itself
conda install pyyaml

# OpenMM Install Before Exafold
# If you are on plain-vanilla x86 machines
conda install -c omnia openmm
# If you have a different hardware, you will need to match
# your CUDA version and hardware to a build of openmm
# or build it yourself in the worst case.
#  -> this one is for OLCF Summit w/ default cuda module
conda install openmm -c omnia-dev/label/cuda101

Install:
If you are on OLCF Summit (or any other PPC64LE machine) we will do some funny install methods. A custom fork of MDTraj will be installed from source, side-by-side with your clone of this repository. This fork gives the necessary MDTraj functionality without any architecture-specific components (so you can't use any geometric calculations essentially). We also will pull OpenMM from a specific target which hard-codes a build against CUDA 10.1, if you need a different PPC64LE build of OpenMM, you will have to modify setup.py accordintly.

git clone https://github.com/jkwoods/exafold
cd exafold
git checkout master

# choose develop option if you want to
# make and test source changes
python setup.py [ install || develop ]

Useful for OpenMM on GPU These need to go into you job script to make GPUs accessible for OpenMM

module load <your cuda module name>
# On summit `module load cuda/10.1.168`
export OPENMM_CUDA_COMPILER=`which nvcc`

Requirements:
Python 3
Packages:

• parse
• OpenMM 7.3+
• MDTraj 1.9.3+

To Run:

1. Run a test:

# from ExaFold top directory
cd tests
python run-test-omm_walker.py

2. (incomplete) Edit my_parameters.yaml with your parameters in a working directory you set up (explaination below) 2.1. Run from structure + restraintlist 2.2. (incomplete) Run from protein sequence

Input to the Program:

1. From structure + restraint list
   • my_parameters.yaml with paths to structure, restraint files
2. From protein sequence
   • my_parameters.yaml with a protein sequence or fasta file

Example API calls
user configuration
exafold.mdsystem.ommsystem
exafold.restraints.reader
exafold.restraints.definitions

Notes

1. See previous repos and migrate relevant stuff

The code is currently under development. A full version will:

• [partial] take a protein sequence for input via a config file
• [partial] create a linear protein structure for this sequence
• [outside] calculate distances used as restraints to help the protein fold correctly
• [partial] calculate distances from secondary structure prediction
• [v1 done] read a set of distance restraints
• [v1 done] build a simulation system and apply restraints
• [nostart] run a swarm of protein folding walkers
• [nostart] prune and move walkers to more effectively fold the protein

Short-term Roadmap
v0.1 restraints are read and applied to an OpenMM system
v0.2 hook the sequence to PDB/restraints upstream
v0.x1 launches HPC job with swarm of folding walkers
v0.x2 prune and move walkers
v0.3 updates to config for more control of workflow