Research map / Probabilistic and scientific modeling / Molecules and proteins
EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site 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 EquiPocket, an E(3)-equivariant GNN over protein atoms and solvent-accessible surface probes with a dense attention output layer, for ligand binding site prediction without voxelization.
Problem
Predicting ligand binding sites on target proteins; CNN voxel-based methods handle irregular shapes poorly, are sensitive to rotations, under-model surface geometry, and fail under protein size shift.
Approach
Replace voxelized 3D CNNs with an E(3)-equivariant GNN that combines local surface-probe geometric descriptors, GAT+EGNN global structure modeling, and equivariant message passing over surface atoms, plus a dense attention output layer that adaptively weights layers by neighbor density.
Key results
- EquiPocket reaches DCC 0.423 on COACH420, 0.337 on HOLO4K and 0.545 on PDBbind2020 versus best prior CNN DeepSurf at 0.386, 0.289 and 0.510.
- Failure rate drops to 0.051 on COACH420 compared with 0.054 for DeepSurf and 0.270 for plain EGNN.
- Ablations show the surface message passing module adds roughly 20% improvement in DCC and DCA over EquiPocket-LG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add surface message passing to EquiPocket-LG (ablation) | EquiPocket · COACH420 | DCC: 0.212(0.016) -> 0.423(0.014) | Helped |
| add surface message passing to EquiPocket-LG (ablation) | EquiPocket · COACH420 | DCA: 0.443(0.011) -> 0.656(0.007) | Helped |
| add global structure modeling to the local geometric module (ablation) | EquiPocket-LG · COACH420 | DCC: 0.070(0.009) -> 0.212(0.016) | Helped |
| add global structure modeling to the local geometric module (ablation) | EquiPocket-LG · COACH420 | DCA: 0.171(0.008) -> 0.443(0.011) | Helped |
| add local geometric modeling to the global structure module (ablation) | EquiPocket-LG · COACH420 | DCC: 0.159(0.016) -> 0.212(0.016) | Helped |
| add local geometric modeling to the global structure module (ablation) | EquiPocket-LG · COACH420 | DCA: 0.373(0.021) -> 0.443(0.011) | Helped |
| use dense attention output instead of no dense attention (ablation) | EquiPocket · proteins with fewer than 3,000 atoms | DCC | Helped |
| add relative direction prediction loss (ablation) | EquiPocket · proteins with fewer than 3,000 atoms | DCC | Helped |
| increase the MSMS probe radius from 1.5 to 2 (ablation) | EquiPocket · binding-site prediction | DCC | Hurt |
| increase surface-EGNN cutoff (ablation) | EquiPocket · binding-site prediction | DCC | Mixed |
| increase surface-EGNN depth (ablation) | EquiPocket · binding-site prediction | DCC | Helped |
| replace the full model with DeepSurf | EquiPocket · COACH420 | DCC: 0.386 -> 0.423 | Helped |
Surprising findings
- Trivially applying equivariant GNNs (EGNN, SchNet) to binding site prediction performs worse than CNN-based and geometric-based methods.
- The local geometric module alone (EquiPocket-L) performs poorly, with negligible performance by itself.
Limitations
- GPU memory grows substantially with surface probe density and cutoff radius.
- Local geometric module alone is insufficient; performance depends on combining modules.
Experimental setup
Datasets: scPDB COACH420 HOLO4K PDBbind v2020
Benchmarks: COACH420 HOLO4K PDBbind2020
Metrics: DCC DCA failure rate success rate
Baselines: Fpocket P2rank DeepSite Kalasanty DeepSurf RecurPocket GAT GCN GCN2 SchNet EGNN
Builds on: EGNN GAT GMN MSMS P2rank
Closest papers by meaning
- HeMeNet: Heterogeneous Multichannel Equivariant Network for Protein Multi-task Learning (AAAI 2025 · similarity 0.87)
- Equivariant Masked Position Prediction for Efficient Molecular Representation (ICLR 2025 · similarity 0.87)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.86)
- Improving Equivariant Graph Neural Networks on Large Geometric Graphs via Virtual Nodes Learning (ICML 2024 · similarity 0.86)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.85)
- Equivariant Spherical Transformer for Efficient Molecular Modeling (NeurIPS 2026 · 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.