Carbohydrate-Protein Interaction Site Prediction

Official repository for CArbohydrate-Protein Site IdentiFier: from the paper Structure-Based Neural Network Protein-Carbohydrate Interaction Predictions at the Residue Level

Citation: Structure-Based Neural Network Protein-Carbohydrate Interaction Predictions at the Residue Level Samuel Canner, Sudhanshu Shanker, Jeffrey Gray

CAPSIF is a deep learning method to determine the carbohydrate-binding residues of proteins given a protein structure. Here we present two CAPSIF models - CAPSIF:Voxel (CAPSIF:V) and CAPSIF:Graph (CAPSIF:G). Both models use convolutions, but different representations of the data. CAPSIF:V uses a voxelized representation and a 3D-UNet architecture to decipher carbohydrate-binding residues. CAPSIF:G uses a graph representation with an Equivariant Graph Neural Network architecture. For further details, check out our paper in Frontiers.

Quick Setup Guide

We suggest using a Micromamba/conda environment for the installation.
Steps:
>> micromamba create -n capsif
>> micromamba install python=3.9 -c conda-forge
>> micromamba install pytorch torchvision cudatoolkit=11.5 -c pytorch -c nvidia -c conda-forge
>> micromamba install biopython pandas colorama scikit-learn matplotlib tqdm py3dmol -c conda-forge
To install pyrosetta, create a ~/.mambarc (.condarc) file with the following content:
--------------------------------------------------------
channels:
- https://USERNAME:PASSWORD@conda.rosettacommons.org
- defaults
--------------------------------------------------------
>> micromamba install pyrosetta

or

get the capsif.yml file from mamba_environment directory
set the usename and password for pyrosetta access.
and run:

>> micromamba create -f capsif.yml -no_rc

Download Weights

First run the commands

mkdir capsif_v/models_DL/
mkdir capsif_g/models_DL/

The weights of each model are stored on our remote server data.graylab.jhu.edu/CAPSIF/

Download my_checkpoint_best_36_2A_CACB_vector_coord_I_clean_data.pth.tar to CAPSIF/capsif_v/models_DL/

Download cb_model.pth.tar to CAPSIF/capsif_g/models_DL/

SYSTEM REQUIREMENTS:

Python3 (3.9)  
PyRosetta4 2023.06+release.27ba97a  py39_0  
biopython==1.81
colorama==0.4.6
IO==0.0.1  
matplotlib==3.7.0  
numpy==1.24.2  
pandas==1.5.3  
pytorch==1.13.1 py3.9_cuda11.7_cudnn8.5.0_0
tqdm==4.64.0  
py3Dmol==1.8.0

PREDICTION:

• For a single PDB you can use the Jupyter Notebook ./sample_notebook.ipynb

• For prediction of multiple PDBs in a directory, we primarily suggest use of ./predict_directory.py or ./notebook_predict_directory.ipynb

• Usage of predict_directory.py

>> python ./predict_directory.py --dir [working_directory/default: 'sample_dir/']
    --v_model [CAPSIF:V Model/default:"./capsif_v/models_DL/my_checkpoint_best_36_2A_CACB_vector_coord_I_clean_data.pth.tar"]
    --g_model [CAPSIF:G Model/default:"./capsif_g/models_DL/cb_model.pth.tar"]
    --out [output_directory/defalt: 'sample_dir']
    --make_pdb [default: True]

Returns: [output_directory]/capsif_predictions.txt with a list of binding residues for each protein and model and [output_directory]/*.pdb with the pdbs with the BFactor identifying the binding Residues

Current settings for B Factor visualization

BFactor = 0.0 : Nonbinder

BFactor = 40.0 : CAPSIF:G Predicted Binder

BFactor = 59.9 : CAPSIF:V Predicted Binder

BFactor = 99.9 : CAPSIF:V and CAPSIF:G Predicted Binder

For conventional command line predictions for each model

• CAPSIF:V - ./capsif_v/README.md for usage of the chimera-integrated script predictor_tool.py

• CAPSIF:G - ./capsif_g/README.md for usage of predict_directory.py

PAPER REPRODUCTION

DATASET INFORMATION

DATASET PREPARATION:

You do not need this step for testing. Go to the training step directly

1. Identify Rosetta readable PDB files using:
   data_preparation/pyrosetta_readable_finding.py

2. Randomly separate PDB files to Train, Test, and Val types.

Use np.random.permutation for random indexing and select segments as per your given ratio. or use:

data_preparation/make_train_and_test_random.py

3. Make simplified pdb data files for faster access during train/test/val using data_preparation/pdb_2_interaction_file_converter.py

TRAINING and TESTING

For each model's reproduction after running data_preparation steps, please refer to each directory's README

CAPSIF_V

CAPSIF:G