Prediction of human-Bacillus anthracis protein-protein interactions using multi-layer neural network

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Predicting host-pathogen protein-protein interactions (PPI) is crucial for understanding disease mechanisms.
• Supervised learning methods for PPI prediction are limited by the scarcity of experimental data.
• Feature selection techniques, such as triplet amino acids, have shown promise in PPI prediction.

Purpose of the Study:

• To compare a neural network model against a Support Vector Machine (SVM) for predicting host-pathogen PPI.
• To evaluate the utility of amino acid quadruplets, pairwise sequence similarity, and human interactome properties as features.
• To enhance the accuracy and performance of PPI prediction models.

Main Methods:

• Implemented neural network and SVM models using Python Sklearn library.
• Utilized a combination of features: amino acid quadruplets, pairwise sequence similarity, and human interactome properties.
• Compared model performance against published predictors and applied them to the human-B.anthracis interaction prediction.

Main Results:

• The neural network model, utilizing quadruplet features and network properties, outperformed the SVM model.
• The neural network model demonstrated significant improvement in overall performance for human-viral PPI prediction compared to triplet-based predictors.
• The model achieved good accuracy in predicting human-B.anthracis PPI, with Gene Ontology analysis revealing functions related to immunology.

Conclusions:

• Neural network models incorporating advanced features like amino acid quadruplets offer superior performance for host-pathogen PPI prediction.
• The developed model provides a valuable tool for studying interactions between hosts and pathogens, particularly in cases with limited experimental data.
• Further improvements in machine learning techniques and feature selection are essential for advancing PPI research.