High throughput molecular docking using Vina-GPU + Docker

This package contains a minimalistic python API for high throughput docking by using VinaGPU via a Docker image.

Features:

1. Can be used to dock on multiple GPUs, multiple workers per GPU.
2. CPU workers using AutoDock Vina python API can be run in parallel to the GPU workers

Installation

1. Pre-requisites:

1. Nvidia driver version>=515.43.04
2. Working nvidia-docker GPU runtime.
3. Python3.8+ conda environment with Rdkit installed

---

2. Install the package, dependencies:

git clone https://github.com/andriusbern/vinaGPU && cd vinaGPU
pip install -e .
pip install meeko docker scipy dimorphite-dl vina

3. Pull the docker image

sudo docker pull andriusbern/vina-gpu:latest

The docker image contains:

• Cuda 11.7
• Vina-GPU (compiled with boost 1.77.0, cuda 11.7, proper OpenCL dependencies)
• Protein preprocessing tools:
  • ADFR Suite
  • pdb_tools

---

Usage: 8 parallel GPU workers (on 4 GPUS) + 8 CPU workers (8 threads each)

import time
from vinagpu import parallel_dock

target_pdb_path = 'examples/P21918.pdb'
output_subfolder = 'test_docking'

with open('examples/valid_smiles.txt', 'r') as f:
    smiles = f.read().splitlines()

t0 = time.time()

parallel_dock(target_pdb_path=target_pdb_path, 
              smiles=smiles,      
              output_subfolder=output_subfolder,
              num_cpu_workers=8, exhaustiveness=8, threads_per_cpu_worker=8, # CPU worker parameters
              gpu_ids=[0,1,2,3], workers_per_gpu=2, search_depth=5)          # GPU Worker parameters

t1 = time.time()
print(f'Docked ligands per second: {len(smiles) / (t1 - t0)}'
print(f'Total time: {t1 - t0}')

Usage, single GPU worker

import os
from vinagpu import VinaGPU

# Docking example on P21918 (DRD5 - D(1B) dopamine receptor)
target_pdb_path = os.path.join('examples', 'P21918.pdb') 
active_site = (2.753, 0.994, -7.633) # Active site coordinates of P21918
output_subfolder = 'P21918_test' # results stored at: "./P21918_test"

smiles = [
    'NC[C@@H](NC(=O)c1ccc(F)cc1)C1CCCC1',
    'CCc1cccc(C(C)(C)NCCc2ccc(OC)c(OC)c2)c1',
    'CCN1CCCC1CNS(=O)(=O)c1ccc(F)cc1',
    'O=C(Cn1cccc1)Nc1ccc(Cl)cc1Cl']

vina_docker = VinaGPU()

scores = vina_docker.dock(
    target_pdb_path=target_pdb_path,
    smiles=smiles,
    output_subfolder=output_subfolder, 
    active_site_coords=active_site,
    verbose=True)