Using Short Molecular Dynamics Simulations to Determine the Important Features of Interactions in Antibody-Protein...

Area of Science:

• Computational biology
• Protein engineering
• Machine learning in drug discovery

Background:

• Machine learning methods for protein design have advanced but often yield inaccurate predictions.
• Distinguishing true protein binding interfaces from false positives remains a computational challenge.

Purpose of the Study:

• To identify critical interaction features for stable protein binding interfaces.
• To enhance the accuracy of computational protein design models.

Main Methods:

• Conducted molecular dynamics simulations on 20 antibody-protein complexes.
• Evaluated persistence, energy, and stability of salt bridges, hydrogen bonds, and hydrophobic interactions.
• Trained and tested random forest classifiers using identified interaction features and established metrics (IE, BSA).

Main Results:

• Only stabilized hydrogen bonds consistently persist and contribute to binding.
• Stabilized interactions (both residues) showed significantly longer persistence and stronger energies.
• Incorporating persistent interaction features reduced false positive rates by 2-5 fold.

Conclusions:

• Stabilized hydrogen bonds are crucial for reliable protein binding.
• The identified persistent interaction features significantly improve the performance of machine learning models for protein binding interface prediction.