[Paper Review] DiffDock-PP: Rigid Protein-Protein Docking with Diffusion Models

Mohamed Amine Ketata, Cedrik Laue+7 more|arXiv (Cornell University)|Apr 8, 2023

Protein Structure and Dynamics26 citations

TL;DR

DiffDock-PP formulates rigid protein-protein docking as a diffusion-generative problem to sample poses and ranks them with a learned confidence model, achieving state-of-the-art performance on DIPS with faster runtimes than many baselines.

ABSTRACT

Understanding how proteins structurally interact is crucial to modern biology, with applications in drug discovery and protein design. Recent machine learning methods have formulated protein-small molecule docking as a generative problem with significant performance boosts over both traditional and deep learning baselines. In this work, we propose a similar approach for rigid protein-protein docking: DiffDock-PP is a diffusion generative model that learns to translate and rotate unbound protein structures into their bound conformations. We achieve state-of-the-art performance on DIPS with a median C-RMSD of 4.85, outperforming all considered baselines. Additionally, DiffDock-PP is faster than all search-based methods and generates reliable confidence estimates for its predictions. Our code is publicly available at $ exttt{https://github.com/ketatam/DiffDock-PP}$

Motivation & Objective

• Motivate rigid-body protein-protein docking as a generative task to capture multi-modal pose distributions.
• Develop a diffusion-based model to map ligand poses (one protein relative to another) on the 6-DOF manifold of rigid motions.
• Leverage SE(3)-equivariant architectures and an intrinsic-diffusion framework compatible with protein symmetry and rigidity.
• Train a confidence model to rank generated poses by likelihood of being near-ground-truth and select the best pose.
• Demonstrate state-of-the-art performance on DIPS and substantial speedups over traditional search-based docking methods.

Proposed method

• Model proteins as residue-level graphs with SE(3)-equivariant score and confidence networks.
• Define diffusion over the product space of translations and 3D rotations to sample ligand poses conditioned on the receptor.
• Use forward diffusion on T(3) and SO(3) with scores in the respective tangent spaces to generate poses.
• Train using denoising score matching on intrinsic manifolds and low-temperature sampling at inference for focused mode concentration.
• Train a separate confidence model to predict whether a sampled pose has L-RMSD below a 5Å threshold, and rank poses by this confidence.
• Output the pose with the highest predicted confidence from the diffusion samples.

Experimental results

Research questions

• RQ1 Can a diffusion generative model effectively approximate the distribution of rigid-body poses for protein-protein docking?
• RQ2 Does sampling multiple poses with a learned confidence-based ranking yield better docked structures than single-shot predictions or traditional docking baselines?
• RQ3 What are the efficiency and accuracy trade-offs of DiffDock-PP compared to state-of-the-art docking methods on DIPS?
• RQ4 Can intrinsic diffusion over the product space of translations and rotations improve generalization for rigid docking tasks?

Key findings

%<2	%<5	%<10	Median	%<2	%<5	%<10	Median	Runtime
34	41	46	11.95	36	42	53	8.60	4.2
42	50	55	4.85	45	52	63	4.23	153
71	79	86	0.67	72	82	91	0.54	153

• On DIPS, DiffDock-PP achieves a median Complex RMSD (C-RMSD) of 4.85 with 40 samples, outperforming all baselines.
• With 40 samples, DiffDock-PP achieves 42% of predictions with C-RMSD < 2Å and 50% with < 5Å for complex RMSD, and 45% and 52% respectively for I-RMSD, with a median I-RMSD of 4.23
• DiffDock-PP is 5 to 60x faster on GPU than common search-based docking software.
• Even with a single sample per complex, DiffDock-PP outperforms most baselines while preserving lower runtimes.
• Oracle-like selection (perfectly picking the best sample) yields substantial upper-bound gains, e.g., 0.67% C-RMSD and 0.54% I-RMSD at the 40-sample oracle setting.
• The model’s performance improves when filtering predictions by the confidence model, indicating effective ranking of proposed poses.