Hybrid Docking

The hybrid docking workflow combines traditional pose generation with GNN rescoring for optimal accuracy. This is the recommended approach for production use.

Why Hybrid Docking?

Traditional docking algorithms are good at:

Generating diverse pose conformations
Sampling the binding site efficiently
Running quickly

But they struggle with:

Accurate affinity ranking
Correlation with experimental data

The GNN excels at:

Predicting binding affinity
Ranking poses correctly
Correlating with experimental pEC50

The hybrid approach combines these strengths.

Workflow

Phase 1: Pose Generation
┌─────────────────────────────────────────┐
│  Hierarchical Docking (Vina scoring)    │
│  → Generate 50+ diverse poses           │
└────────────────────┬────────────────────┘
                     │
Phase 2: GNN Rescoring
┌────────────────────▼────────────────────┐
│  SE(3)-Equivariant GNN                  │
│  → Predict pEC50 for each pose          │
└────────────────────┬────────────────────┘
                     │
Phase 3: Ranking & Output
┌────────────────────▼────────────────────┐
│  Rank by GNN pEC50                      │
│  → Output top-K poses                   │
└─────────────────────────────────────────┘

Basic Usage

pandadock hybrid -r protein.pdb -l ligand.sdf \
                 --center 10 20 30 --box 20 20 20 \
                 -m models/best_model.pt

Full Options

pandadock hybrid \
    --receptor protein.pdb \
    --ligand ligand.sdf \
    --center 10 20 30 \
    --box 20 20 20 \
    --model model.pt \
    --output-dir hybrid_results/ \
    --num-poses 50 \
    --top-k 10 \
    --fast                    # Optional: quick mode

Command Options

Option	Default	Description
`-r/--receptor`	Required	Receptor PDB file
`-l/--ligand`	Required	Ligand file (SDF/MOL2/PDB)
`--center`	Required	Grid center coordinates (X Y Z)
`--box`	Required	Grid box dimensions (X Y Z)
`-m/--model`	Required	GNN model checkpoint
`-o/--output`	hybrid/	Output directory
``-n/–num-poses``\| 50		Poses to generate for rescoring
`--top-k`	10	Top poses to keep after rescoring
`--fast`	False	Use fast mode (fewer poses)

Output Files

The output directory contains:

hybrid_results.csv: Rankings with GNN and Vina scores
pose_1_pec50_X.XX.pdb: Top pose structures
complex_1.pdb: Protein-ligand complexes

Example output table:

Rank  GNN pEC50    GNN Energy    Vina Energy    Activity
1     7.234        -9.87         -8.23          0.92
2     6.891        -9.41         -7.89          0.85
3     6.543        -8.93         -8.45          0.78
...

Performance

Typical timings on GPU:

Phase 1 (50 poses): 10-30 seconds
Phase 2 (rescoring): 1-2 seconds
Total: 15-35 seconds per ligand

Compared to traditional docking alone:

~5x better ranking correlation
Same pose generation quality
Minimal overhead from GNN

Best Practices

Generate many poses: Use --num-poses 50 or more
Keep diverse poses: The GNN will find the best ones
Use trained model: Train on similar targets if possible
Check activity probability: High confidence = reliable prediction

Virtual Screening

For screening many ligands:

for ligand in ligands/*.sdf; do
    pandadock hybrid -r protein.pdb -l "$ligand" \
                     --center 10 20 30 --box 20 20 20 \
                     -m model.pt -o "results/$(basename $ligand .sdf)"
done

Or use the Python API for parallelization.

Comparison with Traditional Docking

Metric	Traditional	Hybrid
Affinity Correlation (R)	0.12	0.67
Ranking Accuracy	Low	High
Speed	Fast	Fast
Pose Quality	Good	Good