Feature selection and classification of protein-protein complexes based on their binding affinities using machine...

Area of Science:

• Computational biology
• Molecular biology
• Bioinformatics

Background:

• Protein-protein interactions (PPIs) are fundamental to cellular functions.
• Accurately predicting the binding affinity of protein complexes remains a significant challenge.
• Understanding binding strength is key for deciphering biological pathways and disease mechanisms.

Purpose of the Study:

• To develop a machine learning model for predicting protein-protein complex binding affinities.
• To identify key sequence features that correlate with binding strength.
• To create an effective tool for analyzing protein-protein interaction networks.

Main Methods:

• Compiled a dataset of 185 heterodimeric protein-protein complexes with experimental binding affinities.
• Extracted 610 sequence-based features from protein complex sequences.
• Employed the Ranker search method (Attribute Evaluator + Ranker) for feature selection.
• Utilized machine learning algorithms, including Support Vector Machines (SVM), to classify complexes into high and low affinity groups.

Main Results:

• Achieved 76.1% accuracy using a combination of nine selected features with SVM via 10-fold cross-validation.
• Attained 83.3% accuracy on an independent test set of 30 protein complexes.
• Demonstrated the effectiveness of sequence features in predicting binding affinity.

Conclusions:

• The developed machine learning approach provides an effective method for predicting protein-protein binding affinity.
• This tool can aid in identifying interacting partners within protein-protein interaction networks.
• The findings are applicable to understanding human-pathogen interactions based on interaction strength.