Research map / Probabilistic and scientific modeling / Molecules and proteins
Group Ligands Docking to Protein Pockets
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 GroupBind, a diffusion-based molecular docking framework that jointly docks groups of ligands sharing a protein pocket, using group-ligand message passing and protein-ligand triangle attention, and reports new SOTA on PDBBind blind docking.
Problem
Molecular docking methods treat each protein-ligand pair in isolation, ignoring that ligands binding the same pocket tend to adopt similar poses.
Approach
Dock a group of ligands to the same pocket simultaneously, passing messages among group ligands (via a protein-independent ligand-ligand graph) and using triangle attention between protein-ligand and group-ligand pair embeddings to enforce interaction consistency; integrated into DiffDock with an auxiliary distance loss and a group-averaged confidence model.
Key results
- GroupBind-P2R-A reaches 48.8% top-5 RMSD<2Å on PDBBind blind docking vs. 41.3% for DiffDock (top-1 2Å: 33.2% for P2R-S vs. 32.4% for DiffDock).
- With the native reference pocket, GroupBind-Ref-A achieves 36.3% top-1 RMSD<2Å and 84.5% top-5 RMSD<5Å, versus 32.4% and 75.6% for DiffDock.
- In the ablation with a smaller model and 10 samples, perfect-selection success rate rises from 30.6% (DiffDock) to 42.9% (full GroupBind).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GroupBind-P2R-S instead of DiffDock | DiffDock · PDBBind v2020 test, Top-5 ligand RMSD | % < 2 Å: 41.3 -> 47.4 | Helped |
| Use GroupBind-P2R-A instead of DiffDock | DiffDock · PDBBind v2020 test, Top-5 ligand RMSD | % < 2 Å: 41.3 -> 48.8 | Helped |
| Use GroupBind-P2R-S instead of DiffDock | DiffDock · PDBBind v2020 test, Top-1 ligand RMSD | 50th percentile RMSD: 3.6 -> 3.4 | Helped |
| Use GroupBind-P2R-A instead of DiffDock | DiffDock · PDBBind v2020 test, Top-1 ligand RMSD | 50th percentile RMSD: 3.6 -> 3.2 | Helped |
| Add message passing across group ligands (ablation) | GroupBind · PDBBind v2020 test, perfect selection with 10 samples | % < 2 Å: 30.6 -> 35.2 | Helped |
| Add triangle attention to MP-GL (ablation) | GroupBind · PDBBind v2020 test, perfect selection with 10 samples | % < 2 Å: 35.2 -> 38.8 | Helped |
| Add auxiliary distance prediction to MP-GL + Att (ablation) | GroupBind · PDBBind v2020 test, perfect selection with 10 samples | % < 2 Å: 38.8 -> 42.9 | Helped |
| Augment with ligands from homologous proteins | GroupBind · New proteins, Top-1 ligand RMSD | % < 2 Å: 20.5 -> 29.6 | Helped |
| Augment with ligands from homologous proteins | GroupBind · New proteins, Top-5 ligand RMSD | % < 2 Å: 34.1 -> 38.6 | Helped |
| Use augmented ligands from the same target group | GroupBind · PDBBind test subsets with available augmented ligands | Docking success rate | Helped |
| Use augmented ligands when no matching ligands are available | GroupBind · PDBBind test subset without group ligands | Docking performance | No clear effect |
| Increase the maximum ligand count from 1 to 2 (ablation) | GroupBind-Ref-S · PDBBind test, perfect selection | % < 2 Å: 53.5 -> 57.9 | Helped |
Surprising findings
- Training the confidence model to directly predict group scores underperformed; averaging individual confidence scores worked better.
- Dissimilar ligands binding the same pocket were found to help more than similar ones (Spearman correlation -0.39 between group similarity and RMSD).
Limitations
- High computational cost of triangle attention limits dense protein-group graphs.
- Requires availability of other ligands binding to the same protein; no benefit for complexes without group ligands.
- Possible increase in unreasonable protein-ligand distances on PoseBusters due to augmented ligands.
Experimental setup
Datasets: PDBBind v2020 PDBBind (homologous protein ligands via FoldSeek)
Benchmarks: PDBBind v2020 blind docking test set (363 complexes) PoseBusters ComBind
Metrics: top-1 ligand RMSD percentiles top-5 ligand RMSD percentiles % RMSD < 2Å % RMSD < 5Å centroid distance top-K success rate
Baselines: GNINA SMINA EquiBind TANKBind DiffDock P2Rank+SMINA P2Rank+GNINA EquiBind+SMINA EquiBind+GNINA
Builds on: DiffDock TANKBind AlphaFold2 e3nn
Closest papers by meaning
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.88)
- Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.86)
- Fast and Accurate Blind Flexible Docking (ICLR 2025 · similarity 0.86)
- Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · similarity 0.85)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.83)
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · 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.