rivesunder/dockrl

Reinforcment learning and evolutionary strategies applied to protein-ligand docking simulation with Autodock Smina.

DockRL

Setup

You'll need PyTorch, NumPy, and a protein-ligand docking simulation program called Smina, which can be downloaded from SourceForge, to use this repository. Smina is a fork of open-source Autodock Vina and has a static executable that needs to be placed in the repositories root folder.

I am using Ubuntu 18 and virtualenv for virtual environment management, but other Linux distributions and/or anaconda should not be too different. To initialize and activate a new virtual environment:

virtualenv /env_path --python=python3
ource /env_path/bin/activate

DockRL uses MPI for training parallelization. In order to use mpip4py, you'll also need to install some dev tools.

sudo apt update && sudo apt install libopenmpi-dev

Next all you need to do is run setup.py to make sure your virtual environment contains dependencies and that python knows where to look for dockrl modules.

python setup.py develop

Usage

Starting a typical training run looks like this

python dockrl/cmaes.py -pi GraphNN -p 64 -g 100 -c 10

-pi designates the policy to use, of which there are currently 3 options. "MLP" is a simple feed forward MLP, "GraphNN" utilizes a self-attentive graph neural network, and "Params" will apply CMA-ES to the scoring function parameters directly. The flags -p, -g, and -c designate the population size, total number of generations, and cpu threads to utilize during CMA-ES training. Note that Smina can also take advantage of parallelism, but the flag for exhaustiveness must be equal to or greater than the number of cpu threads you want smina to use. That being said you may want to balance the number of threads used for CMA-ES and docking to match your machine, as each individual worker will call a seperate instance of Smina. Also remember to add 1 to the total number of CMA-ES workers to account for the mantle process that orchestrates everything.