DeepVS

This is the PyTorch implementation of the DeepVS neural network architecture, which is describe in the paper Boosting Docking-Based Virtual Screening with Deep Learning.

DeepVS is a deep learning approach to improve the identification of active ligands in docking-based virtual screening. DeepVS uses the output of a docking program and learns how to extract relevant features from basic data such as atom and residues types obtained from protein−ligand complexes.

Running the experiment

The train.py file implements the cross-validation experiment reported in the paper. The code should be intuitive. You can run it as follows:

python train.py

Note that in this version of the code we use ReLU and add dropout. These changes made our architeture more robust.

DUD preprocessed data

In order to run the code, you will need our preprocessed vina ouput data.

After downloading the data, unzip it and put the folder dud_vinaout_deepvs in the same directory as train.py

Prerequisites

Python 2.7 Pytorch 0.2.0_4

Paper Reference

If this code is useful for you somehow please cite our paper:

@article{doi:10.1021/acs.jcim.6b00355,
author = {Pereira, Janaina Cruz and Caffarena, Ernesto Raúl and dos Santos, Cicero Nogueira},
title = {Boosting Docking-Based Virtual Screening with Deep Learning},
journal = {Journal of Chemical Information and Modeling},
volume = {56},
number = {12},
pages = {2495-2506},
year = {2016},
doi = {10.1021/acs.jcim.6b00355},
note ={PMID: 28024405},
URL = {https://doi.org/10.1021/acs.jcim.6b00355},
eprint = {https://doi.org/10.1021/acs.jcim.6b00355}
}

License