Research map / Probabilistic and scientific modeling / Molecules and proteins
Deep Confident Steps to New Pockets: Strategies for Docking Generalization
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.
Introduces the DockGen benchmark for blind-docking generalization across ECOD protein domains, scales DiffDock data and model size with vdM-inspired synthetic data to get DiffDock-L, and proposes Confidence Bootstrapping, a self-training scheme that fine-tunes the diffusion model using confidence-model feedback on unseen protein clusters.
Problem
Blind molecular docking methods based on ML generalize poorly to protein binding pockets unseen in training, and existing benchmarks (PDBBind time split, sequence-similarity splits) leak pocket similarity between train and test sets.
Approach
Define splits by ECOD binding-domain clusters rather than sequence similarity; improve generalization by scaling data/model size and adding synthetic van der Mer-inspired sidechain-as-ligand complexes; and fine-tune a diffusion docking model on unseen clusters by bootstrapping high-confidence samples from a confidence model into the early diffusion steps.
Key results
- DiffDock-L raises ML-docking success on DockGen-full from 7.1% to 22.6%, above the best search-based GNINA (ex. 64) at 17.5%.
- Confidence Bootstrapping raises DiffDock-S on DockGen-clusters from 9.8% to 24.0% success rate, with above 30% in half of the clusters.
- DiffDock-L reaches 50% RMSD<2Å on PoseBusters, versus 38% for DiffDock and 42% for RosettaFold-AllAtom among blind methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Scale training data and model size and add van der Mer-inspired synthetic training examples | DiffDock · DockGen | Top-1 success rate (%<2Å): 7.1% to 22.6% | Helped |
| Fine-tune on protein clusters using Confidence Bootstrapping feedback | DiffDock-S · DockGen-clusters | Top-1 success rate (%<2Å): 9.8% to 24.0% | Helped |
| Increase the training data with MOAD complexes from already-seen protein domains | DiffDock · DockGen | Generalization performance | Helped |
| Increase model size | DiffDock · DockGen | Generalization performance | Helped |
| Add van der Mer-inspired synthetic docking augmentation | DiffDock · DockGen | Generalization performance | Helped |
Surprising findings
- Regression-based ML docking methods (EquiBind, TankBind) score near 0% on DockGen-full and DockGen-clusters despite strong PDBBind time-split results.
- Search-based methods like SMINA and GNINA degrade less than ML methods on DockGen, though even with high exhaustiveness they still miss the true pose in most cases.
Limitations
- Confidence Bootstrapping is evaluated on a small test set of 85 complexes across 8 clusters, with results averaged over two runs.
- Synthetic vdM complexes have unknown affinity and limited ligand chemistry (only the 20 amino acids as ligands).
- Gains depend on the initial accuracy of the score model; clusters with zero initial accuracy did not improve.
Experimental setup
Datasets: PDBBind Binding MOAD BindingDB ProteinMPNN training set PoseBusters
Benchmarks: PDBBind DockGen-full DockGen-clusters PoseBusters
Metrics: top-1 RMSD < 2 Å success rate median RMSD runtime (s)
Baselines: SMINA GNINA P2Rank+GNINA EquiBind TankBind DiffDock RosettaFold-AllAtom
Builds on: DiffDock AlphaFold2 ProteinMPNN ESM
Closest papers by meaning
- DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking (ICLR 2023 · similarity 0.89)
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.89)
- Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.88)
- Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.87)
- Fast and Accurate Blind Flexible Docking (ICLR 2025 · similarity 0.86)
- Group Ligands Docking to Protein Pockets (ICLR 2025 · similarity 0.85)
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.