Deep Local Analysis evaluates protein docking conformations with locally oriented cubes.

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning in structural biology

Background:

• Advances in protein 3D structure prediction highlight the growing importance of understanding protein interactions.
• Accurate determination of protein complex conformations is crucial for biological research.

Purpose of the Study:

• To develop and evaluate a deep learning framework for discriminating near-native protein complex conformations from incorrect ones.
• To leverage local environments around interfacial residues for improved protein interaction analysis.

Main Methods:

• Deep Local Analysis (DLA)-Ranker framework utilizes 3D convolutions on locally oriented cubes representing protein interfaces.
• The method explicitly considers local geometry, neighboring atoms, and solvent accessibility of interfacial residues.
• Performance was assessed on three docking benchmarks comprising half a million conformations.

Main Results:

• DLA-Ranker successfully identifies near-native protein complex conformations from molecular docking ensembles.
• The framework demonstrates competitive or superior performance compared to existing deep learning-based scoring functions.
• The study showcases DLA-Ranker's utility in discovering alternative protein-protein interfaces.

Conclusions:

• Deep Local Analysis (DLA)-Ranker provides an effective deep learning approach for protein complex structure evaluation.
• The method's focus on local interface environments enhances the accuracy of protein interaction prediction.
• DLA-Ranker offers a valuable tool for structural biologists and computational chemists.