DeepFragLib: Fragment Library Construction Software by Deep Neural Network for Ab Initio Protein Structure Prediction
*********************************************************************************************************************

Current Version: 1.0
By Tong Wang ----- March 2019


DEPENDENCIES
************
DeepFragLib could be run with Python2.7 or Python3. 

Some packages should be installed with the specific version:
1. Tensorflow-gpu v1.8.0
2. Numpy v1.15.4

The local version of DeepFragLib does not require any other software to be executed.

However, DeepFragLib requires 4 input files:

1. TARGET.fasta: The fasta sequence of the target protein.

2. TARGET.ss: The 3-class predicted secondary structure by PSIPRED v4.01. 
Users could get the file by PSIPRED online server at https://bioinf.cs.ucl.ac.uk/psipred_new/ and need to remove the first two heading lines. 
Alternatively, a local version of PSIPRED could be downloaded at https://bioinfadmin.cs.ucl.ac.uk/downloads/psipred/. 
Please rename the file as TARGET.ss 

3. TARGET.spd3: The predicted torsion angles by SPOT1D. 
Users could get the file by SPOT1D online server at https://sparks-lab.org/jack/server/SPOT-1D/ and reformat the file by "scripts/reformat_spot1d_server.py". 
Alternatively, a local version of SPOT1D could be downloaded at https://sparks-lab.org/jack/download/ index.php?Download=SPOT1D_local.tgz and reformat the file by "scripts/reformat_spot1d_local.py".


4. TARGET.mat: The predicted residues contact map by SPOT-Contact. 
Users could get the file by SPOT-Contact online server at https://sparks-lab.org/jack/server/SPOT-Contact/ and reformat the file by "scripts/reformat_spot2d_server.py". 
Alternatively, a local version of SPOT-Contact could be downloaded at https://sparks-lab.org/jack/download/index.php?Download=SPOT-Contact_local.tgz and reformat the file by "scripts/reformat_spot2d_local.py".

In addition, if users want to calculate RMSD for template fragment, a PDB file only containing C-alpha coordinates of the target protein should be provided.

Users should put all input files in the directory "examples/inputs/" and ensure that these files have the same formats of the corresponding example files we provided.


EXAMPLE USAGE
*************

An example is in "data/example". To generate the fragment library, type the command as follows:

$ python main.py example

After a period of time, two result files will be generated: example.lib recording detailed information during computation and example.frag recording dihedral angles and coordinates with the same format of the output file by NNMake in Rosetta.


CONFIGURATION
*************

Users could change parameters, such as the input file directory, the output file directory, the choice of exclusion of homologous proteins and the choice of RMSD calcualtion in "main.py".

In addition, users could choose to use either GPU or CPU to run the software by modifying "gpu_id" in "main.py". Notably, using GPU can greatly accelerate the computational speed since the deep neural networks are sophisticated.


OTHER INFORMATION
*****************

Users could use our web server at https://structpred.life.tsinghua.edu.cn/DeepFragLib.html as an alternative choice.

REFERENCES
**********
1, Tong Wang#, Yanhua Qiao, Wenze Ding, Wenzhi Mao, Yaoqi Zhou* and Haipeng Gong*,“Improving fragment sampling for ab initio protein structure prediction using deep neural networks ”, Nature Machine Intelligence, 1: 347-355, 2019.

2, Tong Wang#, Yuedong Yang, Yaoqi Zhou, and Haipeng Gong*, “LRFragLib: an effective algorithm to identify fragments for de novo protein structure prediction”, Bioinformatics, 33(5): 677-683, 2017.


TROUBLESHOOTING
***************

Please contact [email protected] for any problems users may encounter using DeepFragLib.