gcndesign 2.2.1

Neural network model for predicting amino-acid sequence from a protein backbone structure

GCNdesign

A neural network model for prediction of amino-acid probability from a protein backbone structure.

Built with

• pytorch
• numpy
• pandas
• tqdm

Getting Started

Install

To install gcndesgn through pip

pip install gcndesign

Usage

Quick usage as a python module

from gcndesign.prediction import Predictor

gcndes = Predictor(device='cpu') # 'cuda' can also be applied
gcndes.pred(pdb='pdb-file-path') # returns list of amino-acid probabilities

Usage of scripts

gcndesign_predict.py

To predict amino-acid probabilities for each residue-site

gcndesign_predict.py  YOUR_BACKBONE_STR.pdb

gcndesign_autodesign.py

To design 20 sequences in a completely automatic fashion

gcndesign_autodesign.py  YOUR_BACKBONE_STR.pdb  -n 20

For more detailed usage, please run the following command

gcndesign_autodesign.py -h

Note

The gcndesign_autodesign script requires pyrosetta software. Installation & use of pyrosetta must be in accordance with their license.

External Packages

• gcndesign_autodesign.py: PyRosetta

Documents

• Method summary

Note

A critical issue has fixed and the parameters were re-trained on a new dataset (CATH v4.3 S40 dataset). This change has stabilized the prediction, but has not been reflected in the document above. So there are inaccuracies in the description and figures.

Dataset

The dataset used for training GCNdesign is available here

• dataset.tar.gz: Training/T500/TS50 dataset
• dataset_cath40.tar.bz2: CATH-v4.3 S40 dataset (used for the latest parameter training)

Lisence

Distributed under MIT license.

Acknowledgments

The author was supported by Grant-in-Aid for JSPS Research Fellows (PD, 17J02339). Koga Laboratory of Institutes for Molecular Science (NINS, Japan) has provided a part of the computational resources. Koya Sakuma (yakomaxa) gave a critical idea for neural net architecture design in a lot of deep discussions. Naoya Kobayashi (naokob) created excellent applications to help broader needs, ColabGCNdesign and FolditStandalone_Sequence_Design.