Research map / Probabilistic and scientific modeling / Molecules and proteins
Fast and Accurate Blind Flexible Docking
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 FABFlex, a regression-based multi-task model that jointly predicts pocket sites and holo ligand and pocket structures from apo inputs for fast blind flexible docking on PDBBind v2020.
Problem
Blind flexible molecular docking requires predicting pocket location and holo ligand/protein structures from apo inputs, but existing flexible methods rely on slow diffusion sampling and rigid methods ignore protein flexibility.
Approach
A single-pass, end-to-end regression model with three FABind-layer modules (pocket prediction, ligand docking, pocket docking) linked by an iterative coordinate update mechanism, trained with a two-stage pretraining curriculum and partial teacher forcing.
Key results
- FABFlex achieves 40.59% ligand RMSD < 2Å on all test cases versus 33.00% for DynamicBind and 36.75% for DiffDock-L.
- On unseen receptors FABFlex reaches 32.46% ligand RMSD < 2Å, above all competitors (all below 30%).
- Average inference of 0.49 s per pair, about 208x faster than DynamicBind (102.12 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use separate ligand and pocket docking modules rather than one module for both structures (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 27.06 -> 40.59 | Helped |
| Add iterative update between ligand and pocket docking modules (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 19.80 -> 40.59 | Helped |
| Exchange predictions between modules instead of updating each module only internally (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 35.31 -> 40.59 | Helped |
| Use integrated pocket prediction instead of P2Rank (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 34.22 -> 40.59 | Helped |
| Increase iterative updates from 1 to 2 (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 21.45 -> 36.96 | Helped |
| Increase iterative updates from 2 to 4 (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 36.96 -> 39.27 | Helped |
| Increase iterative updates from 4 to 6 (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 39.27 -> 40.59 | Helped |
| Increase iterative updates from 6 to 8 (ablation) | FABFlex · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 40.59 -> 38.94 | Hurt |
| Use FABFlex rather than DynamicBind | PDBBind v2020 test, all cases · PDBBind v2020 test, all cases | Average runtime (s): 102.12 -> 0.49 | Helped |
| Use FABFlex rather than the strongest listed baseline, DiffDock-L | PDBBind v2020 test, all cases · PDBBind v2020 test, all cases | Ligand RMSD < 2 Å (%): 36.75 -> 40.59 | Helped |
| Use FABFlex rather than DynamicBind | PDBBind v2020 flexible test · PDBBind v2020 flexible test | PB-valid & Ligand RMSD < 2 Å (%): 14.57 -> 13.91 | Hurt |
Surprising findings
- On PoseBusters physical validity, FABFlex's PB-valid & RMSD<2Å (13.91%) is below DiffDock-L (16.89%), showing diffusion models produce more physically valid poses.
- Pocket RMSD of FABFlex (1.10 Å mean) is worse than DynamicBind on all cases, and it does not outperform DynamicBind in pocket refinement.
Limitations
- Pocket RMSD is worse than DynamicBind on the full test set, so pocket refinement is weaker than diffusion-based flexible docking.
- Physical validity (PoseBusters) lags diffusion methods; many validity checks such as minimum distance to protein remain low.
- Evaluated only on PDBBind v2020 with AlphaFold2 apo structures, and the limited unseen-receptor set (114 complexes) constrains generalization claims.
Experimental setup
Datasets: PDBBind v2020
Benchmarks: PDBBind v2020 test set (303 complexes) PoseBusters test suite
Metrics: ligand RMSD pocket RMSD % ligand RMSD < 2Å pocket center MAE/RMSE/EucDist classification accuracy runtime
Baselines: Vina Glide Gnina TankBind FABind FABind+ DiffDock DiffDock-L DynamicBind
Builds on: FABind FABind+ EGNN ESM-2 AlphaFold2
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)
- Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · similarity 0.86)
- AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.86)
- Group Ligands Docking to Protein Pockets (ICLR 2025 · similarity 0.86)
- Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design (ICML 2024 · similarity 0.84)
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.