IIDL-PepPI

Protein complex structural data is growing at an unprecedented pace, but its complexity and diversity pose significant challenges for protein function research. Although deep learning models have been widely used to capture the syntactic structure, word semantics, or semantic meanings of polypeptide and protein sequences, these models often overlook the complex contextual information of sequences. Here, we propose IIDL-PepPI, a deep learning model designed to tackle these challenges by employing data-driven and interpretable pragmatic analysis to profile peptide-protein interactions (PepPIs). IIDL-PepPI constructs bidirectional attention modules to represent the contextual information of peptides and proteins, enabling pragmatic analysis. It then adopts a progressive transfer learning framework to simultaneously predict PepPIs and identify binding residues in specific interactions, providing a solution for multi-level in-depth profiling.

Given the complexity and instability of individuals in configuring the environment, we strongly recommend that users use IIDL-PepPI 's online prediction Web server, which can be accessed through http://bliulab.net/IIDL-PepPI/.

Fig. 1: Data preparation workflow and network architecture of IIDL-PepPI. a Data preparation workflow of IIDL-PepPI, in which the public databases used include RCSB PDB, PDBe, and UniProt. b Network architecture of IIDL-PepPI for peptide-protein binary interaction prediction and binding residue recognition, including sequence representation, feature encoding, bi-attentional module, and decoding. Based on the biological sequence pragmatic analysis, the bi-attention module explicitly integrates features from the peptide and protein sides to distinguish different peptide-protein-specific interactions. c The progressive transfer learning architecture. The initial stage of IIDL-PepPI commences with pre-training peptide-protein binary interactions using sequence-level datasets and the coarse-grained learning of basic network parameters. Subsequently, in the second phase, we transfer the parameters of the basic network, replace the decoder, and conduct fine-grained fine-tuning of the model using residue-level dataset for precise prediction of peptide- and protein-binding residues in specific peptide-protein pairs.

1 Installation

1.1 Create conda environment

conda create -n iidl python=3.10
conda activate iidl

1.2 Requirements

The main dependencies used in this project are as follows (for more information, please see the environment.yaml file):

python  3.10
biopython 1.81
huggingface-hub 0.19.4
numpy 1.26.2
pandas 2.1.3
scikit-learn 1.3.2
scipy 1.11.4
tokenizers 0.15.0
torch 2.1.1+cu118
torchaudio 2.1.1+cu118
tqdm 4.66.1
transformers 4.35.2

Note If you have an available GPU, the accelerated IIDL-PepPI can be used to predict peptide-protein binary interactions and pair-specific binding residues. Change the URL below to reflect your version of the cuda toolkit (cu118 for cuda=11.6 and cuda 11.8, cu121 for cuda 12.1). However, do not provide a number greater than your installed cuda toolkit version!

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118

For more information on other cuda versions, see the pytorch installation documentation.

1.3 Tools

Two multiple sequence alignment tools and three databases are required:

SCRATCH-1D 1.2
IUPred2A \
ncbi-blast 2.13.0
ProtBERT \

Databases:

nrdb90 (http://bliulab.net/sAMPpred-GAT/static/download/nrdb90.tar.gz)

nrdb90: We have supplied the nrdb90 databases on our webserver. You need to put it into the utils/ directoy and decompress it.

Note that all the defalut paths of the tools and databases are shown in config.yaml. You can change the paths of the tools and databases by configuring config.yaml as you need.

SCRATCH-1D, IUPred2A, ncbi-blast, and ProtBERT are recommended to be configured as the system envirenment path. Your can follow these steps to install them:

1.3.1 How to install SCRATCH-1D

Download (For linux, about 6.3GB. More information, please see https://download.igb.uci.edu/)

wget https://download.igb.uci.edu/SCRATCH-1D_1.2.tar.gz
tar -xvzf SCRATCH-1D_1.2.tar.gz

Install

cd SCRATCH-1D_1.2
perl install.pl

Note: The 32 bit linux version of blast is provided by default in the 'pkg' sub-folder of the package but can, probably should, and in some cases has to be replaced by the 64 bit or Mac OS version of the blast software for improved performances and compatibility on such systems.

Finally, test the installation of SCRATCH-1D

cd <INSTALL_DIR>/doc
../bin/run_SCRATCH-1D_predictors.sh test.fasta test.out 4

Note: If your computer has less than 4 cores, replace 4 by 1 in the command line above.

1.3.2 How to install IUPred2A

For download and installation of IUPred2A, please refer to https://iupred2a.elte.hu/download_new. It should be noted that this automation service is only applicable to academic users. For business users, please contact the original authors for authorization.

After obtaining the IUPred2A software package, decompress it.

tar -xvzf iupred2a.tar.gz

Finally, test the installation of IUPred2A

cd <INSTALL_DIR>
python3 iupred2a P53_HUMAN.seq long

1.3.3 How to install ncbi-blast

Download (For x64-linux, about 220M. More information, please see https://blast.ncbi.nlm.nih.gov/doc/blast-help/downloadblastdata.html)

wget https://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/2.13.0/ncbi-blast-2.13.0+-x64-linux.tar.gz
tar -xvzf ncbi-blast-2.13.0+-x64-linux.tar.gz

Add the path to system envirenment in ~/.bashrc.

export BLAST_HOME={your_path}/ncbi-blast-2.13.0+
export PATH=$PATH:$BLAST_HOME/bin

Finally, reload the system envirenment and check the ncbi-blast command:

source ~/.bashrc
psiblast -h

Note: The purpose of IIDL-PepPI with the help of ncbi-blast is to extract the position-specific scoring matrix (PSSM). It should be noted that for sequences that cannot be effectively aligned, the PSSM is further extracted by blosum62 (which can be found in utils/blosum62.txt).

1.3.4 How to install ProtBERT

Download and install (More information, please see https://huggingface.co/Rostlab/prot_bert or https://github.com/agemagician/ProtTrans)

wget https://zenodo.org/records/4633691/files/prot_bert.zip

or

git lfs install
git clone https://huggingface.co/Rostlab/prot_bert

1.4 Inatsll IIDL-PepPI

To install from the development branch run

git clone git@github.com:ShutaoChen97/IIDL-PepPI.git
cd IIDL-PepPI/utils
tar -xvzf nrdb90.tar.gz
cd ..

Besides, due to the limit of 2G file size uploaded by Git LFS, the comparison file used by IIDL-PepPI to reduce the dimensionality of pre-training (ProtBERT) features is available through our IIDL-PepPI online Web server (about 3G).

wget http://bliulab.net/IIDL-PepPI/static/download/protein_webserver.pkl
mv protein_webserver.pkl saved_models/protbert_feature_before_pca/

Finally, configure the Defalut path of the above tool and the database in config.yaml. You can change the path of the tool and database by configuring config.yaml as needed.

2 Usage

It takes 2 steps to predict peptide-protein binary interaction and peptide-protein-specific binding residues:

(1) Replace the default peptide sequence in the example/example_peptide_ 1.fasta file with your peptide sequence (FASTA format). Similarly, replace the default protein sequence in the example/example_protein_ 1.fasta file with your protein sequence (FASTA format). If you don't want to do this, you can also test your own peptide-protein pairs by modifying the two sequence file paths passed in by the test.sh script (the two parameters are -pep_fasta for peptide and -pro_fasta for protein, respectively).

(2) Then, run test.sh to make multi-level prediction, including binary interaction prediction and combined residue recognition. It should be noted that test.sh automatically calls the scripts generate_peptide_features.py, generate_protein_features.py, and generate_pssm.py to generate the multi-source isomerization characteristics of peptides and proteins.

bash test.sh

Note you can running python test.py -h to learn the meaning of each parameter.

If you want to retrain based on your private dataset, find the original IIDL-PepPI model in model/IIDL-PepPI.py. The IIDL-PepPI source code we wrote is based on the Pytorch implementation and can be easily imported by instantiating it.

3 Problem feedback

If you have questions on how to use IIDL-PepPI (or PDB-BRE), feel free to raise questions in the discussions section. If you identify any potential bugs, feel free to raise them in the issuetracker.

In addition, if you have any further questions about IIDL-PepPI, please feel free to contact us [stchen@bliulab.net or shutao.chen@bit.edu.cn]

4 Citation

If you find our work useful, please cite us at

@article{chen2023pdb,
  title={PDB-BRE: A ligand--protein interaction binding residue extractor based on Protein Data Bank},
  author={Shutao Chen, Ke Yan, and Bin Liu},
  journal={Proteins: Structure, Function, and Bioinformatics},
  year={2023},
  publisher={Wiley Online Library}
}

@article{chen2024protein,
  title={Protein language pragmatic analysis and progressive transfer learning for profiling peptide-protein interactions},
  author={Shutao Chen, Ke Yan, Xuelong Li, and Bin Liu},
  journal={submitted},
  year={2024},
  publisher={}
}