Research map / Probabilistic and scientific modeling / Molecules and proteins
Rigid Protein-Protein Docking via Equivariant Elliptic-Paraboloid Interface Prediction
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.
Proposes ElliDock, which predicts elliptic paraboloid binding interfaces for both proteins and derives the rigid docking transform by aligning them, using an SE(3)-equivariant graph network (EPIT) for fast docking.
Problem
Rigid protein-protein docking: predicting the SE(3) transformation that maps unbound protein structures to their docked pose, quickly and without steric clashes, with strong performance especially on antibody-antigen complexes.
Approach
Instead of regressing pocket point-cloud registration (EquiDock) or diffusion sampling (DiffDock-PP), the model predicts a pair of elliptic paraboloid interfaces, one per protein, and computes the docking transform in closed form by making the two interfaces coincide; an overlap loss keeps the proteins on opposite sides of the interface to avoid clashes.
Key results
- On DB5.5, ElliDock reaches median IRMSD 11.134 vs. DiffDock-PP 15.060 and EquiDock 11.971, with median CRMSD 12.995 (best among learned methods) and inference time 36.7 vs. 2103.1 s for DiffDock-PP.
- On SAbDab/RAbD, ElliDock achieves median CRMSD 11.541 and IRMSD 11.319, outperforming EquiDock and DiffDock-PP on IRMSD while DiffDock-PP has a slightly better mean CRMSD (12.560 vs. 13.402).
- ElliDock is the fastest method evaluated (91.2 s on SAbDab test versus 8308.7 s for DiffDock-PP), but HDock and AlphaFold-Multimer score higher in DockQ and RMSD on most metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add L_fit (ablation) | ElliDock · DB5.5 | IRMSD median: 13.02 -> 11.13 | Helped |
| Add L_overlap (ablation) | ElliDock · DB5.5 | CRMSD median: 13.95 -> 13.00 | Helped |
| Add L_ref (ablation) | ElliDock · DB5.5 | IRMSD median: 12.49 -> 11.13 | Helped |
| Add L_fit and L_overlap (ablation) | ElliDock · DB5.5 | IRMSD median: 13.24 -> 11.13 | Helped |
| Add L_fit and L_ref (ablation) | ElliDock · DB5.5 | IRMSD median: 13.94 -> 11.13 | Helped |
| Add L_overlap and L_ref (ablation) | ElliDock · DB5.5 | CRMSD median: 14.48 -> 13.00 | Helped |
| Add rotation refinement Q_r (ablation) | ElliDock · DB5.5 | IRMSD median: 13.22 -> 11.13 | Helped |
| Use ElliDock instead of DiffDock-PP | rigid docking · DB5.5 test | inference time: 2103.1 -> 36.7 | Helped |
| Use ElliDock instead of DiffDock-PP | rigid docking · SAbDab test (RAbD) | CRMSD mean: 12.560 -> 13.402 | Hurt |
| Use ElliDock instead of DiffDock-PP | rigid docking · SAbDab test (RAbD) | IRMSD mean: 12.401 -> 11.550 | Helped |
Surprising findings
- Pre-trained HDock achieves much lower median RMSD and higher DockQ than all learned methods, which the authors attribute possibly to data leakage.
- Removing the overlap loss raises CRMSD but lowers IRMSD in the ablation (IRMSD 12.35 vs. 11.13 full), so the components do not uniformly help.
Limitations
- Targets rigid docking only and ignores conformational changes of the proteins.
- Uses soft constraints, so predicted results may not fully satisfy the intended geometric properties.
- Results are weaker than HDock and AlphaFold-Multimer on several metrics, and ElliDock is only competitive with, not clearly better than, DiffDock-PP on CRMSD for SAbDab.
Experimental setup
Datasets: DB5.5 SAbDab RAbD
Benchmarks: DB5.5 test set RAbD (SAbDab test set)
Metrics: CRMSD IRMSD DockQ inference time
Baselines: HDock AlphaFold-Multimer EquiDock DiffDock-PP
Builds on: PaiNN EquiDock DiffDock-PP Kabsch algorithm
Closest papers by meaning
- Equivariant Scalar Fields for Molecular Docking with Fast Fourier Transforms (ICLR 2024 · similarity 0.85)
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.84)
- Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.84)
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.83)
- Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · similarity 0.83)
- Towards Stable Representations for Protein Interface Prediction (NeurIPS 2024 · similarity 0.83)
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.