Unified Sampling and Ranking for Protein Docking With DFMDock

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Artificial Intelligence in Drug Discovery

Background:

• Current diffusion models for protein-protein docking often separate structure generation from decoy ranking.
• This separation can lead to suboptimal performance and increased computational complexity.
• There is a need for unified models that integrate both processes efficiently.

Purpose of the Study:

• To introduce DFMDock, a unified diffusion model for protein-protein docking.
• To integrate generative sampling and energy-based ranking within a single framework.
• To improve docking accuracy and decoy ranking efficiency using physically motivated supervision.

Main Methods:

• DFMDock employs a diffusion model trained with force matching and energy contrastive objectives.
• It predicts denoising forces to guide the reverse diffusion process for structure generation.
• It predicts a scalar energy for decoy ranking, eliminating the need for separate confidence models.

Main Results:

• DFMDock achieved a 4.9% Top-1 and 31.6% Oracle success rate on the Docking Benchmark 5, outperforming DiffDock-PP.
• The model generalizes to unseen targets without requiring multiple sequence alignments (MSAs).
• The learned energy function surpassed Rosetta energy and model-derived scores in decoy ranking, creating enriched energy landscapes.

Conclusions:

• DFMDock presents an efficient and physically grounded approach to diffusion-based protein docking.
• The unified model successfully integrates generation and ranking, improving performance.
• DFMDock offers a promising alternative to existing methods, particularly for targets lacking MSAs.