Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Co-folding models like AlphaFold excel at protein complex prediction but struggle with targets lacking multiple sequence alignments (MSAs).
• Structure-based protein-protein docking offers an alternative by predicting bound structures from unbound monomers without MSAs.

Purpose of the Study:

• To adapt co-folding models for structure-based protein-protein docking.
• To develop a generative docking model named AF2Dock by modifying AlphaFold-Multimer (AF-M).

Main Methods:

• Replaced the template module of AF-M with a docking module.
• Trained the modified model end-to-end using a flow-matching objective.
• Applied the method to the OpenFold implementation of AF-M, creating AF2Dock.

Main Results:

• AF2Dock demonstrated competitive performance on the PINDER-AF2 benchmark and antibody/nanobody datasets using non-holo inputs.
• AF2Dock outperformed all tested docking methods for nanobody complexes.
• AF2Dock provided orthogonal predictions, successfully solving cases where co-folding models failed.

Conclusions:

• AF2Dock represents a novel adaptation of co-folding models for structure-based protein-protein docking.
• Full-parameter fine-tuning is crucial for AF-M component performance in docking.
• The study highlights the potential and limitations of integrating co-folding architectures into docking strategies.