CAMP: a Convolutional Attention-based Neural Network for Multi-level Peptide-protein Interaction Prediction

CAMP is a sequence-based deep learning framework for multifaceted prediction of peptide-protein interactions, including not only binary peptide-protein interactions, but also corresponding peptide binding residues.

Notice (Updated in 2022/10/19)

Since CAMP exploits a series of sequence-based features, you CANNOT use CAMP when the primary sequence information of peptides and proteins is unknown. In addition, you need to replace any non-standard amino acid with 'X' first.

Requirement

Trained and tested on a linux server with GeForce GTX 1080 and the running environment is as follows: Python2.7, Keras=2.0.8, Tensorflow=1.2.1, RDKit (conda install -y -c conda-forge rdkit), CUDA (GPU computation)

Running CAMP

1. To predict with our example code, first create a directory example_prediction and unzip the feature dictionaries by unzip example_data_feature.zip and use the test data with the command python -u predict_CAMP.py. Options are:
   -b: The batch size, default: 256.
-l: The padding length for peptides, default: 50.
-f: The fraction of GPU memory to process the model, default: 0.9.
-m: The prediction model, 1 indicates binary prediction task and 2 indicates dual predictions, default: 1.
-p: The The padding length for proteins p, default: 800.
2. To generate features, please refer to the instructuion in "Data curation".
3. We recently added run.sh for users with no ML backgrounds to easiley test the model. Just simply run command sh run.sh.

Data curation

Here we offer the protocol to construct the benchmark dataset. Due to the copyright issues, we can not provide the complete benchmark dataset. You may follow the procedures in ./data_prepare/ to reproduce the benchmark dataset and corresponding features (or features of your own data). For cluster-based cross validations, after finishing all the following procedures, you can get alignment results from https://github.com/mengyao/Complete-Striped-Smith-Waterman-Library and use the scripts in ./cluster/ to obtain the similarity matrix and split clusters of peptides and proteins based on the similarity matrix.

• STEP 0: Downloading data from three sources

1. RCSB PDB : Download the fasta files from ftp://ftp.wwpdb.org/pub/pdb/derived_data/pdb_seqres.txt.gz and pdb files

2. UniProt : Download pdb_chain_uniprot.tsv.gz containing the UniProt IDs of proteins in PDB complexes, which will be used to map the UniProt sequences later from https://www.ebi.ac.uk/pdbe/docs/sifts/quick.html

3. DrugBank : Download full database.xml from DrugBank, to access the data, you need to create an account first. (requiring licenses) Then only keep DITs that belong to "Peptides" Category.

• STEP 1: Process and filter PDB data using the functions in ./data_prepare/step1_pdb_process.py

• STEP 2: Generate labels of peptide binding residues of PepPIs from PDB by the functions and detailed procedures in ./data_prepare/step2_pepBDB_pep_bindingsites.py

• STEP 3: Generated sequnece-based features Combine all data from DrugBank and PDB, then shuffle pairs to obtain negative samples. After that, you need to generate the corresponding features (you can either use the online servers or download the softwares):

  1. Secondary structure : http://scratch.proteomics.ics.uci.edu/explanation.html#SSpro IS NOT CORRECT. PLEASE DOWNLOAD THE LINUX 2018 VERSION HERE: https://download.igb.uci.edu/ (SCRATCH-1D release 1.2 (2018, linux version, 6.3 GB))
  2. Intrinsic Disorder : https://iupred2a.elte.hu/
  3. PSSM matrix: ftp://ftp.cnbi.nlm.nih.gov/blast/executables/blast+/LATEST/
  4. Use the functions in ./data_prepare/step3_generate_features.py to process raw output files. Then put all feature dicts in ./dense_feature_dict/

• STEP 4: Proprecess data with feature dictionaries

1. format the peptide-protein data like (protein sequence, peptide sequence, protein_ss, peptide_ss) and generate a test data file called "test_filename" for
2. Use the command python -u preprocess_features.py test_filename to obtain processed feature dicts in ./preprocessing/

Recent Update 2022/09

Some researchers find that the web server generating secondary structures http://scratch.proteomics.ics.uci.edu/explanation.html#SSpro would generate different results which we used to train and evaluate CAMP by installing the 2018 LINUX version. Such differences may due to the updated predicting algorithm of SSPro in 2020. We investigated such variance and found that around 5-10% residues would be predicted into distinct SS-classes. Since our model was trained using the 2018-linux verion software, please generate these features by using the LINUX 2018 version (SCRATCH-1D release 1.2 (2018, linux version)) for inference. We'll update our model with new features soon.

Recent Update 2022/10

For better inference, here we provided Keras model files and corresponding feature generation scripts with 20 standard AA + "U/B/O/Z" + "X" for unknown AA. Pay attention to that if you want to train with your own data or make inference.

Since many researchers request for training CAMP on their own data, I re-implemented CAMP using Python3+pytorch recently and the model file with corresponding training/predicting scripts will be uploaded this month. ( You can still use the original version trained by Python2+Keras).

License

This software is copyrighted by Machine Learning and Computational Biology Group @ Tsinghua University.

The algorithm and data can be used only for NON COMMERCIAL purposes.