Research map / Probabilistic and scientific modeling / Molecules and proteins

DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein interface and docking prediction · 2 citations (OpenAlex)

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 DeltaDock, a two-stage docking framework that reframes pocket prediction as contrastive pocket-ligand alignment and uses a bi-level coarse-to-fine E(3)-equivariant refinement with physics-informed losses and structure correction for accurate, physically valid docking.

Problem

Geometric deep learning docking methods handle only blind or only site-specific settings, ignore protein side chains, struggle with large pockets, and often produce physically invalid poses.

Approach

Treat blind-docking pocket prediction as contrastive alignment between ligand and candidate pocket embeddings (selecting from an ensemble of DSDP and P2Rank pockets), then refine poses with a bi-level residue-then-atom E(3)-equivariant network initialized by GPU sampling, trained with vdW and bound-matrix penalties and followed by torsion alignment and SMINA minimization.

Key results

• In blind docking on the PDBbind time split, DeltaDock reaches 47.4% RMSD<2 Å success versus 36.0% for DiffDock (about 31% relative improvement).
• On the timesplit unseen set, DeltaDock achieves 40.8% success, 20.1 points above FABind.
• Blind docking takes about 2.97 seconds per protein-ligand pair on average.

Measured effects reported in the paper

Surprising findings

• Ablating the structure sampling initialization drops RMSD<2 Å success from 47.4% to 16.0% on PDBbind, showing initialization dominates performance.
• Training with the physics-based losses and structure correction slightly lowers RMSD accuracy while improving physical validity.

Limitations

• Rigid protein assumption: only ligand flexibility is modeled.
• Pocket selection relies on candidate pockets from DSDP and P2Rank; other pocket predictors hurt training due to non-druggable hard negatives.

Experimental setup

Datasets: PDBbind v2020 PoseBusters

Benchmarks: PDBbind time split test PDBbind timesplit unseen PoseBusters

Metrics: RMSD success rate below 2 Å centroid distance PoseBusters validity checks DCC VCR

Baselines: DiffDock FABind TANKBind EquiBind DSDP VINA SMINA GNINA

Builds on: DSDP DiffDock Attentive-FP GVP SMINA

Closest papers by meaning

• Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge (ICML 2024 · similarity 0.91)
• Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · similarity 0.89)
• Group Ligands Docking to Protein Pockets (ICLR 2025 · similarity 0.88)
• Fast and Accurate Blind Flexible Docking (ICLR 2025 · similarity 0.88)
• Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion (ICML 2024 · similarity 0.87)
• AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.86)