Characterizing informative sequence descriptors and predicting binding affinities of heterodimeric protein complexes

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Protein-protein interactions (PPIs) are vital for biological processes, influencing therapeutics and protein engineering.
• Current machine learning methods often rely on structure and functional data for PPI binding affinity prediction.
• Predicting binding affinity solely from amino acid sequences presents a significant challenge.

Purpose of the Study:

• To develop a sequence-based method for predicting the binding affinity of heterodimeric protein complexes.
• To identify key sequence descriptors that are most informative for binding affinity classification.
• To improve upon existing machine learning algorithms for PPI binding affinity prediction.

Main Methods:

• A support vector machine (SVM) based classifier, SVM-BAC, was developed.
• Optimal feature selection identified 14 informative sequence descriptors from physicochemical, energetic, and conformational properties.
• The method was trained and tested on heterodimeric protein complexes to classify low and high binding affinity.

Main Results:

• SVM-BAC achieved high accuracy in classifying binding affinity, with training accuracy of 85.80% and test accuracy of 83.33%.
• The model demonstrated superior performance compared to existing machine learning algorithms.
• Analysis revealed that apparent partition energy, PCA IV, and beta turn frequency are crucial for prediction.

Conclusions:

• The sequence-based SVM-BAC method effectively classifies and predicts binding affinity using 14 selected features.
• High binding affinity complexes exhibit higher average numbers of beta turns and hydrogen bonds at interfaces.
• This approach offers a novel way to predict binding affinity without structural information.