Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge

Yufei Huang, Odin Zhang, Lirong Wu, Cheng Tan, Haitao Lin, Zhangyang Gao, Siyuan Li, Stan. Z. Li

TL;DR

Re-Dock addresses the lack of realistic pocket-sidechain flexibility in molecular docking by formulating docking as a diffusion-bridge problem on geometric manifolds. It introduces an energy-to-geometry mapping inspired by Newton-Euler mechanics to inject interaction priors into geometry-based diffusion bridges, and co-models binding energy with poses using an SE(3)-equivariant framework. The approach enables autoregressive sidechain updates and explicit energy guidance, achieving superior performance on apo-dock, flexible redocking, cross-dock, and PoseBuster benchmarks, while remaining efficient for large-scale screening. This work advances practical docking by capturing induced-fit dynamics without requiring holo-pocket inputs, with potential impact on drug design and protein engineering pipelines.

Abstract

Accurate prediction of protein-ligand binding structures, a task known as molecular docking is crucial for drug design but remains challenging. While deep learning has shown promise, existing methods often depend on holo-protein structures (docked, and not accessible in realistic tasks) or neglect pocket sidechain conformations, leading to limited practical utility and unrealistic conformation predictions. To fill these gaps, we introduce an under-explored task, named flexible docking to predict poses of ligand and pocket sidechains simultaneously and introduce Re-Dock, a novel diffusion bridge generative model extended to geometric manifolds. Specifically, we propose energy-to-geometry mapping inspired by the Newton-Euler equation to co-model the binding energy and conformations for reflecting the energy-constrained docking generative process. Comprehensive experiments on designed benchmark datasets including apo-dock and cross-dock demonstrate our model's superior effectiveness and efficiency over current methods.

Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge

TL;DR

Re-Dock addresses the lack of realistic pocket-sidechain flexibility in molecular docking by formulating docking as a diffusion-bridge problem on geometric manifolds. It introduces an energy-to-geometry mapping inspired by Newton-Euler mechanics to inject interaction priors into geometry-based diffusion bridges, and co-models binding energy with poses using an SE(3)-equivariant framework. The approach enables autoregressive sidechain updates and explicit energy guidance, achieving superior performance on apo-dock, flexible redocking, cross-dock, and PoseBuster benchmarks, while remaining efficient for large-scale screening. This work advances practical docking by capturing induced-fit dynamics without requiring holo-pocket inputs, with potential impact on drug design and protein engineering pipelines.

Abstract

Accurate prediction of protein-ligand binding structures, a task known as molecular docking is crucial for drug design but remains challenging. While deep learning has shown promise, existing methods often depend on holo-protein structures (docked, and not accessible in realistic tasks) or neglect pocket sidechain conformations, leading to limited practical utility and unrealistic conformation predictions. To fill these gaps, we introduce an under-explored task, named flexible docking to predict poses of ligand and pocket sidechains simultaneously and introduce Re-Dock, a novel diffusion bridge generative model extended to geometric manifolds. Specifically, we propose energy-to-geometry mapping inspired by the Newton-Euler equation to co-model the binding energy and conformations for reflecting the energy-constrained docking generative process. Comprehensive experiments on designed benchmark datasets including apo-dock and cross-dock demonstrate our model's superior effectiveness and efficiency over current methods.
Paper Structure (26 sections, 9 equations, 4 figures, 6 tables, 1 algorithm)

This paper contains 26 sections, 9 equations, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Backgrounds
  4. Diffusion Bridge for Non-geometric domains
  5. Problem Statement
  6. Method
  7. Diffusion Prior Bridge on Geometries
  8. Injecting interaction prior into Geometry Bridges.
  9. learning bridges by co-modeling energy and poses
  10. Experiments
  11. Experimental Setups
  12. Results on flexible redocking and apo docking (Q1)
  13. Quality of sidechain pose generation (Q2)
  14. Performace comparison on cross-dock (Q3)
  15. Ablation Study (Q4)
  16. ...and 11 more sections

Figures (4)

  • Figure 1: The illustration of our motivation. (a) The dotted line represents current protocols take docked pockets as input, which are not accessible in realistic scenarios and provide hints or leakage for ligand poses prediction. (b) The result of DiffDock (pdb_id: 6nsv); the steric clashes are highlighted with red circles, where the ligand overlaps with the protein surface (i.e., sidechains).
  • Figure 2: The illustration of Re-Dock Framework. We aim to simulate the induced fitting process with geometric prior bridges. Our key designs are threefold: ① The pocket sidechains displace the most flexibility for inducing interactions. Thus, we generate the sidechain conformations (the blue and purple sticks are the conformations of two steps before and after, respectively; we omit other sidechains for simplicity) via torsion angle updates while docking. ② We explore a novel generative model, the geometric prior bridge for reflecting the energy-constrained fitting process. Compared with the diffusion processes (the red curves), the prior bridge process (the blue line) is augmented with problem-dependent prior and thus more fast and accurate to generate. ③ For explicit modeling of interaction and constructing prior bridges over geometries, we propose an energy-to-geometry mapping module inspired by Newton-Euler equations.
  • Figure 3: The illustration of sidechain updates. (a) Up to four sidechain $\theta^{sc}$ angles (formally, $\chi$ angles) have a sequential order. (b) Rotating $\theta^{sc}_1$ will affect the coordinates of atoms in $\theta^{sc}_2$, $\theta^{sc}_3$ and $\theta^{sc}_4$. It's similar for rotating $\theta^{sc}_2$ and $\theta^{sc}_3$. The atom groups of the later angles will accumulate noise from the former angles which complicates the latter's denoising process.
  • Figure 4: Case Studies on Plasmin gallus1976prevention, a real-world protein target and its inhibitor ligand SFTI-1 de2021sunflower. The numbers (e.g. HIS-57) indicate the name and id of residues. Re-Dock replicates the bound crystal conformations for ligands (Docked) and sidechains (Key Change) on the apo (Top, PDB: 1QRZ) and predicted (Bottom) structure of plasmin.