MetaScore: A Novel Machine-Learning-Based Approach to Improve Traditional Scoring Functions for Scoring...

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Protein-protein interactions are crucial for biological processes, necessitating knowledge of complex 3D structures.
• Computational docking is a vital tool for determining these structures, but accurately scoring docked models remains a challenge.

Purpose of the Study:

• To develop a novel machine learning-based approach, MetaScore, to improve the scoring of docked protein complex conformations.
• To enhance the identification of near-native models from docking simulations.

Main Methods:

• Developed MetaScore, a random forest classifier trained on protein-protein interfacial features.
• Included physicochemical properties, energy terms, geometric properties, evolutionary conservation, and traditional scoring function scores as features.
• Scored docked conformations by averaging the MetaScore (RF classifier) and traditional scoring function scores.

Main Results:

• MetaScore consistently outperformed nine traditional scoring functions in success and hit rates for top-ranked conformations.
• An ensemble method, MetaScore-Ensemble, combining MetaScore variants, further improved performance over individual MetaScore variants.

Conclusions:

• Machine learning can significantly improve protein-protein docking scoring by utilizing interfacial features.
• Ensemble methods offer a powerful strategy to combine multiple scoring functions for enhanced prediction accuracy.