QSL_CPI

This repo provides the codes used in the study on Quasi-Supervised Learning Algorithm (QSL) for Compound-Protein Interaction Prediction. Please check the following link for a more detailed explanation of this work.

• Quasi-Supervised Strategies for Compound-Protein Interaction Prediction

Our study is based on the machine learning strategy presented in the following paper:

• Quasi-supervised learning for biomedical data analysis

If you use this repo in your work, please cite these two papers given above.

Highlights of our method:

• Compound-Protein pairs at hand are divided into two datasets:
  • C1 consists of the true postive samples. (experimentally validated interactions)
  • C0 consists of unlabeled samples. (compound-protein pairs that do not have a documented interaction)
• QSL calculates posterior probability of each pair belonging to C0 and C1.
• Kolmogorov-Smirnov Method detects the threshold that draws the boundary of the overlap between the C1 and C0.
• QSL can predict the interaction profiles of unlabeled pairs accurately without requiring true negative compound-protein pairs.
• QSL is robust against data imbalance between true positives and unlabeled compound-protein pairs.
• QSL operates on similarity structure between protein and compound pairs directly without requiring a feature vector representation.

Installation

git clone https://github.com/caki35/qsl_cpi.git
cd qsl_cpi

Istallation using conda:

conda create --name qsl_cpi python=3.8
conda activate qsl_cpi
conda install -c conda-forge rdkit
pip install -r requirements.txt

Download Data

We provide a script scripts/download_data.py for retrieving either compound or protein data from KEGG databases. You need just give a txt file including id list of the data that you want to download.

python scripts/download_data.py Data/Proteins/GPCR/kegg_id.txt protein

This script saves amino acid sequences of all given protein KEGG ids into a unified file named protein.fasta.

python scripts/download_data.py Data/Compounds/GPCR/kegg_id.txt compound 

This script saves the structure information of all given compound KEGG ids into kegg_mol folder in .mol file format.

Generating SMILES strings

You can convert mol files of compounds into SMILES representations using scripts/smiles_converter.sh
It requires molconverter console program of JCHEM. The program defines SMILES by following Daylight’s SMILES specification rules. You must install this program and provide its sysytem path into scripts/smiles_converter.sh

bash scripts/smiles_converter.sh MOL_FOLDER_PATH

All .mol files in given folder are converted into SMILES representation and saved into folder smiles_output.

Similarity Calculation

To use QSL for CPI prediction, you must construct similarity matrices for both compound and protein data first. We provide various options to that end. If you have your own similarity matrices, you can pass this step.

Compound-Compound Similarity Measurement

SMILES Kernels
To calculate similarity scores between SMILES pairs, run following script with folder path including SMILES files. You must choose one of the kernel options:

• 'NLCS' : Normalized Longest Common Subsequence (NLCS)
• 'CLCS' : Combination of Longest Common Subsequence Models (CLCS)
• 'LINGO3': LINGO-3 Similarity
• 'LINGO4': LINGO-4 Similarity
• 'LINGO5': LINGO-5 Similarity
• 'LINGO4TF': LINGO-4 Based Term Frequency (TF) Cosine Similarity
• 'LINGO4TFIDF': Frequency-Inverse Document Frequency (TF-IDF) Cosine Similarity

python kernels/compound_smiles_kernels.py SMILES_FOLDER_PATH LINGO4TF

Graph Kernel
To calculate similarity scores between compounds using their graph representations, run following script with folder path including .mol files.

python kernels/compound_graph_kernel.py Data/Compounds/GPCR/GPCR_kegg_mol

Molecular-Fingerprints Kernels
To calculate similarity scores between compounds using their fingerprints, run following script with folder path including .mol files. You must choose one of the fingerprint options:

• 'KCFS' : KEGG Chemical Function and Substructures
• 'ECFP4': Extended-Connectivity Fingerprints
• 'FCFP4': FunctionalClass Fingerprintsx
• 'MACCS': Molecular ACCess System

python kernels/compound_fingerprint_kernels.py MOL_FOLDER_PATH ECFP4

NOTE: To use KCFS, you must install KCF-Convoy.

The results will be saved into Similarity_Matrices in .csv file format.

Protein-Protein Similarity Measurement

Run the following script to construct a similarity matrix for proteins using Smith-Waterman Algorithm. You need to input.fasta file that contains amino acid sequences of all proteins.

python kernels/protein_seq_kernel.py FASTA_FILE_PATH

Run QSL algorithm

python QSL.py CONFIG_FILE_PATH

The detailed of config.yml:
save_dir:The directory in which to save the output of the QSL algorithm.
compound_similarity_matrix: The path of similarity matrix for compounds.
protein_similarity_matrix: The path of similarity matrix for proteins.
interaction_info: The path of .txt file including interaction information at hand.
n_range: The range of n which is the main parameter that QSL optimizes.
grid_search_optimizer: if True, the algorithm performs grid search during finding the best n instead of considering each n in range.
single_n_flag: if True, QSL algorithm runs at a certain n instead of optimizing it.
trivial_names: if True, the KEGG IDs in results will be converted into their trivial names.

qsl:
  save_dir: results_TFIDF_NR #the directory to which the results are saved
  compound_similarity_matrix: Similarity_Matrices/Nuclear_Receptor/compound_smiles_LINGO4TFIDF.csv
  protein_similarity_matrix: Similarity_Matrices/Nuclear_Receptor/protein_NSWA_nr.csv
  interaction_info: Data/interaction_data_nr.txt
  n_range: [1,50]
  grid_search_optimizer: False
  single_n:
    single_n_flag: False
    n: 21
  trivial_names: True

The result folder will contain 4 files:

• The graph of cost function with which n is optimized.
• Posterior probability distributions of compound-protein pairs to belong to the true positive dataset and their Kolmogorov-Smirnov Analysis.
• The list of posterior probabilities of all pairs.
• The list of predicted interactions.

Cost function when grid search is opened.

Cost function when grid search is closed.