Feature-based classification of native and non-native protein-protein interactions: Comparing supervised and...

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Advances in structural biology yield numerous PPI structures and models.
• Distinguishing physiological PPIs from experimental artifacts or inaccurate models is challenging.

Purpose of the Study:

• To develop computational methods for classifying physiological vs. crystal-packing PPIs.
• To classify near-native vs. inaccurate protein docking models.
• To evaluate the effectiveness of machine learning approaches in PPI analysis.

Main Methods:

• Defined universal interface features for PPIs.
• Employed Support Vector Machines (SVM) for classification.
• Utilized Transductive Support Vector Machines (TSVM) for semi-supervised learning on docking models.

Main Results:

• SVM classifier achieved 93% accuracy in distinguishing experimental PPIs.
• SVM and TSVM classifiers reached 78.9% and 80.3% accuracy on protein docking benchmark, respectively.
• Model re-ranking in protein docking remains a difficult problem.

Conclusions:

• Machine learning, particularly SVM and TSVM, can effectively classify protein-protein interactions and docking models.
• Accurate classification of PPIs is vital for understanding cellular processes.
• Further improvements in protein docking methodologies are needed for reliable model assessment.