ContactGAN

ContactGAN takes a predicted protein contact map as input and outputs a new contact map that better captures the residue-residue contacts

Copyright (C) 2020 Sai Raghavendra Maddhuri, Aashish Jain, Yuki Kagaya, Genki Terashi, Daisuke Kihara, and Purdue University.

License: GPL v3 for academic use. (For commercial use, please contact us for different licensing)

Contact: Daisuke Kihara (dkihara@purdue.edu)

Citation:

@article{sai2021contactgan,
  title={Protein Contact Map Refinement for Improving Structure Prediction Using Generative Adversarial Networks},
  author={Sai Raghavendra Maddhuri Venkata Subramaniya, Genki Terashi, Aashish Jain, Yuki Kagaya, Daisuke Kihara},
  journal={Bioinformatics}
}

About ContactGAN

ContactGAN is a novel contact map denoising and refinement method based on Generative adversarial networks.
ContactGAN can be trained and combined with any contact prediction method to improve and correct similar patterns of errors that creep into the method. Currently ContactGAN is trained and works with 4 contact prediction methods - CCMPred, DeepContact, DeepCov and trRosetta

Pre-required software

Python 3 : https://www.python.org/downloads/  
pytorch : pip/conda install pytorch  
CCMPred : A freely available software. It can be downloaded and installed from here : https://github.com/soedinglab/CCMpred  
DeepContact : A freely available software. It can be downloaded and installed from here : https://github.com/largelymfs/deepcontact   
DeepCov : A freely available software. It can be downloaded and installed from here : https://github.com/psipred/DeepCov  

Instructions

Generate an input contact map file using a method of your choice from the 4 methods described.

ContactGAN Usage

If you are testing with a single method as 1-channel input, run ContactGAN as follows:

python test/denoising_gan_test.py --input=<INPUT contact prediction directory> --G_res_blocks=3 --D_res_blocks=3 --G_path=model/<1-channel directory>/Generator --D_path=model/<1-channel directory>/Discriminator
  --input               Input Contact Map  Directory  
  --G_res_blocks        Number of ResNet blocks in Generator (Default : 3)
  --D_res_blocks        Number of ResNet blocks in Discriminator (Default : 3)
  --G_path              Specify path of Generator model
  --D_path              Specify path of Discriminator model
  

If you are testing with two methods as 2-channel input, run ContactGAN as follows:

python test/denoising_gan2_test.py --input <INPUT contact prediction directory 1> <INPUT contact prediction directory 2> --G_res_blocks=3 --D_res_blocks=3 --G_path=model/<2-channel directory>/Generator --D_path=model/<2-channel directory>/Discriminator
 

If you are testing with three methods as 3-channel input, run ContactGAN as follows:

python test/denoising_gan3_test.py --input <INPUT contact prediction directory 1> <INPUT contact prediction directory 2> <INPUT contact prediction directory 3> --G_res_blocks=3 --D_res_blocks=3 --G_path=model/<3-channel directory>/Generator --D_path=model/<3-channel directory>/Discriminator
 

Output interpretation

Generated output contact map file is the denoised version of the input map.Output file looks exactly same as input file structure-wise.
An example contact map can be found at the bottom of this page.

Visualization

python util/plot_cmap.py --input=<OUTPUT Contact Prediction File>
  --input               Input Contact Map    
  

Tutorial:

Single-Channel
In this tutorial, you'll learn to test ContactGAN for single channel inputs.
For the purpose of this tutorial, please refer to example contact map input and output are provided in data/example_files/

ContactGAN Usage

To run ContactGAN, you will need an input contact map from one of the following 4 methods - CCMpred, DeepCov, DeepContact, or trRosetta.
An example contact map can be found here.
Model files required to run ContactGAN can be found here
Once you have a contact map e.g. CCMpred, you can run ContactGAN as follows:

1. Specify input map directory to --input argument
2. G_res_blocks - Number of Generator ResNet blocks. Specify 6 for trRosetta and 3 for others.
3. D_res_blocks - Number of Disciminator ResNet blocks. Specify 3.
4. G_path - Generator Model Path. If you're using CCMpred you can use this path
5. D_path - Discriminator Model Path. If you're using CCMpred you can use this path

python test/denoising_gan_test.py --input=data/example_files/input/single_channel --G_res_blocks=3 --D_res_blocks=3 --G_path=model/CCMPred/Generator --D_path=model/CCMPred/Discriminator

Multi-Channel (2 channels)
In this tutorial, you'll learn to test ContactGAN for multi-channel (2) inputs.
For the purpose of this tutorial, please refer to example contact map input and output are provided in data/example_files/

ContactGAN Usage

To run ContactGAN with multiple channels i.e. either 2 or 3, you will need specify multiple contact maps as inputs.
An example for multiple channels can be found here.
Model files required to run ContactGAN can be found here

1. Specify input map directories i.e., channel1 and channel2 to --input argument.
2. G_res_blocks - Number of Generator ResNet blocks. Specify 6 for trRosetta and 3 for others.
3. D_res_blocks - Number of Disciminator ResNet blocks. Specify 3.
4. G_path - Generator Model Path. If you're using CCMpred you can use this path
5. D_path - Discriminator Model Path. If you're using CCMpred you can use this path

python test/denoising_gan2_test.py --input data/example_files/input/multi_channel/channel1 data/example_files/input/multi_channel/channel2 --G_res_blocks=3 --D_res_blocks=3 --G_path=model/CCMPred_DeepContact/Generator --D_path=model/CCMPred_DeepContact/Discriminator

Output contact map Visualization

python util/plot_cmap.py --input=data/example_files/output/5OHQA.npy
 

Below is an example visualization for contact maps of CCMpred before and after ContactGAN for protein with PDB ID: 5OHQA.