ProteinDock: A physics-informed layer to improve protein-protein docking reliability

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Interaction Modeling

Background:

• Computational modeling offers insights into protein-protein interactions but struggles with flexible antibody-antigen interfaces.
• Existing methods lack reliability for modeling proteins with unique features like antibodies, which utilize polar-rich loops for antigen binding.

Purpose of the Study:

• To develop and validate ProteinDock, a physics-based tool to enhance the reliability of protein-protein docking.
• To demonstrate ProteinDock's utility in improving antibody-antigen interface predictions.
• To present a computationally efficient strategy for improving protein-protein prediction from sequence data using ProteinDock.

Main Methods:

• Integrated ProteinDock with Rosetta for docking unbound experimentally determined protein structures.
• Evaluated ProteinDock's performance on the Docking Benchmark Set 5.5.
• Utilized a truncated version of ProteinDock to select optimal predictions from multiple deep learning models for sequence-based predictions.

Main Results:

• ProteinDock achieved CAPRI acceptable-quality or better for 80.2% of targets, a significant improvement over vanilla Rosetta (47.4%).
• The truncated ProteinDock strategy proved computationally efficient for selecting optimal predictions from deep learning models.
• A graphical user interface for ProteinDock was developed, facilitating its application.

Conclusions:

• ProteinDock substantially improves the reliability of protein-protein docking, particularly for challenging antibody-antigen interactions.
• The tool offers a viable and efficient alternative to traditional methods and deep learning seed enhancement for sequence-based predictions.
• ProteinDock is accessible and provides a valuable resource for the structural biology and computational chemistry communities.