Official repository for Protein interaction of CArbohydrate Predictor (PiCAP) and CArbohydrate-Protein Site IdentiFier2 (CAPSIF2) from the paper:
Predictions from Deep Learning Propose Substantial Protein-Carbohydrate Interplay
Citation
Predictions from Deep Learning Propose Substantial Protein-Carbohydrate InterplaySamuel W Canner, Ronald L Schnaar, Jeffrey J Gray
bioRxiv 2025.03.07.641884; doi: https://doi.org/10.1101/2025.03.07.641884
Protein interaction of CArbohydrate Predictor (PiCAP) is deep learning model to predict whether a protein binds to a carbohydrate using a protein’s sequence and structure in tandem. PiCAP uses ESM2 embeddings and the input structure to identify if the protein has the capacity to bind to carbohydrates.
Carbohydrate Protein Site IdentiFier 2 (CAPSIF2) is a deep learning model to predict protein-carobhydrate binding sites. CAPSIF2 predicts on a residue level which residues are implicated in protein binding. It uses a 3D equivariant graph neural network for predictions.
All data for the proteome analysis is available at : datasets/proteome_results.xlsx.
The PiCAP protocol requires set of tools to run. To streamline this process, we provide a Docker container that includes all necessary applications pre-installed, offering a straightforward command-line interface for running the PiCAP protocol.
Please note that the provided Docker image includes PyRosetta which require a commercial license for non-academic use. For more details, please refer to: RosettaCommons
To run the application, use the picap or capsif-2 scripts with same same command line options as run_both.py (see local install section below).
mkdir ./input_pdb
mkdir ./output_data
# copy your input files into input_pdb dir
docker run -it -v ./input_pdb:/picap/input_pdb -v ./output_data:/picap/output_data rosettacommons/rosetta:picap --high_plddt --plddt_cutoff 70For a full list of available options, run: docker run -it rosettacommons/rosetta:picap picap --help
mkdir pre_pdb
mkdir output_data
mkdir models_DL
conda env create -f picap.yml
conda activate picap
cd models_DL
wget https://data.graylab.jhu.edu/picap_capsif2/model-picap.pt
wget https://data.graylab.jhu.edu/picap_capsif2/model-capsif2.pt
cd ..
Or you can manually download with the following:
The weights of each model are stored on our remote server data.graylab.jhu.edu/picap_capsif2/
Download model-picap.pt and model-capsif2.pt to capsif2_clean/models_DL/
Put all PDB (or CIF) files into the input_pdb/ directory
python run_both.py
python run_both.py --picap_only
python run_both.py --capsif2_only
python run_both.py --high_plddt --plddt_cutoff 70
plddt_cutoff can be changed to any value, the publication uses 70 as the cutoff for AF2 structures.
the predictions will then be outputted to output_data/predictions_prot.tsv and output_data/predictions_res.tsv for PiCAP and CAPSIF2, respectively.
If running both, then the data will be outputted to output_data/all_predictions.tsv
All predictions for CAPSIF2 are also outputted individually as PDB files in the output_data/ directory.
Put all PDB (or CIF) files into the input_pdb/ directory
Load the sample_notebook.ipynb to run a single structure through (no high_plddt option provided), which allows quick analysis and viewing of a single structure.
Load the notebook_predict_directory.ipynb and run the script to predict for all structures in the input_pdb/ directory.
We include the NoCAP and DR datasets in the datasets/ directory with a list of PDBs. All non-RCSB retrievable structures (e.g. designed non-binders and ProGen lysozymes) are at the remote server data.graylab.jhu.edu/picap_capsif2/.
Use case note, due to pyrosetta problems, we only can use PDB files for input so all cif files are converted using Bio.PDB to pdb files and then output as a pdb to the output_data/ directory.
A simplified version of the training code is provided in ./training_code/ for quick modification and alteration if desired.