Research map / Probabilistic and scientific modeling / Molecules and proteins
Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Re-Dock introduces flexible docking (ligand plus pocket sidechain poses) and a geometric diffusion bridge model with Newton-Euler energy-to-geometry mapping that co-models binding energy and poses.
Problem
Molecular docking that relies on holo protein structures leaks information and ignores pocket sidechain flexibility, producing unrealistic poses with steric clashes; realistic flexible docking (apo and cross-dock) remains under-addressed.
Approach
Extends diffusion bridges to geometric manifolds (translation, SO(3) rotation, torsion angles) and injects interaction-energy priors via a Newton-Euler-inspired energy-to-geometry mapping, with autoregressive sidechain torsion generation and a learnable energy head to co-model energy and poses.
Key results
- On PDBBind holo redocking, Re-Dock (40 samples) reaches 53.9% top-1 RMSD<2 Å and median 1.8 Å, vs. 51.8% and 2.0 Å for DiffDock (pocket, 40).
- On cross-dock, Re-Dock (10 samples) reaches 36.1% top-1 RMSD<2 Å and median 2.6 Å, above GNINA-Flex (19.6%) and SMINA-Flex (26.2%).
- On PoseBusters, Re-Dock (40) achieves 50.7% top-1 RMSD<2 Å and 32.8% when also PB-Valid, versus 38.0% and 12.0% for DiffDock (40).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sidechain torsion sampling at inference (ablation) | Re-Dock · PDBBind test, flexible redocking | %<2 Top-1 RMSD: 39.8 -> 53.9 | Helped |
| add sidechain torsion sampling at inference (ablation) | Re-Dock · PDBBind test, flexible redocking | Med. Top-1 RMSD: 3.3 -> 1.8 | Helped |
| add prior energy drift during training (ablation) | Re-Dock · PDBBind test, flexible redocking | %<2 Top-1 RMSD: 50.3 -> 53.9 | Helped |
| add prior energy drift during training (ablation) | Re-Dock · PDBBind test, flexible redocking | Med. Top-1 RMSD: 2.4 -> 1.8 | Helped |
| use learned prior bridge rather than classifier-guided sampling (ablation) | Re-Dock vs DiffDock (pocket) · PDBBind test, flexible redocking | %<2 Top-1 RMSD: 52.7 -> 53.9 | Helped |
| reduce sampling steps to 10 (ablation) | Re-Dock · PDBBind test, flexible redocking | %<2 Top-1 RMSD: 53.9 -> 52.1 | Hurt |
| reduce sampling steps to 10 (ablation) | Re-Dock · PDBBind test, flexible redocking | Med. Top-1 RMSD: 1.8 -> 2.0 | Hurt |
| use Re-Dock instead of DiffDock (pocket) | PDBBind test holo crystal docking · PDBBind test, holo crystal, Top-1 RMSD | %<2: 51.8 -> 53.9 | Helped |
| use Re-Dock instead of DiffDock (pocket) | PDBBind test apo ESMFold docking · PDBBind test, apo ESMFold, Top-1 RMSD | %<2: 37.8 -> 42.9 | Helped |
| use Re-Dock instead of SMINA-Flex | CrossDock · CrossDock, Top-5 RMSD | %<2: 26.2 -> 36.1 | Helped |
Surprising findings
- Removing the prior energy term lowers top-1 RMSD<2 Å from 53.9% to 50.3% on the ablation, and DiffDock with classifier guidance (52.7%) still trails the bridge-trained model.
Limitations
- Pocket prediction and end-to-end blind docking are not explored; p2rank-based results are notably weaker.
- Torsion-space prior drift on SO(2)^M is omitted, and autoregressive sidechain update with that prior is left to future work.
- Training takes about 7 days and the cross-dock benchmark reports only 10 samples per complex for compute reasons.
Experimental setup
Datasets: PDBBind v2020 PoseBusters DUD-E ESMFold predicted apo structures PDB apo structures
Benchmarks: PDBBind test (flexible redocking) ESMFold apo structures apo crystal structures cross-dock (7 DUD-E targets) PoseBusters
Metrics: top-1 RMSD < 2 percent top-1 RMSD < 5 percent median RMSD top-5 RMSD SC-RMSD average runtime PB-Valid pass rate
Baselines: SMINA GNINA LeDock SurfLex QVina GNINA-Flex SMINA-Flex FlexPose DiffDock
Builds on: DiffDock diffusion bridge (prior bridge) Brownian bridge Amber force field
Closest papers by meaning
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.91)
- Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.88)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.87)
- Fast and Accurate Blind Flexible Docking (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.