PandaDock - Molecular Docking with GNN Scoring

SE(3)-Equivariant GNN Scoring for Molecular Docking

Installation | Quick Start | Documentation | Benchmark | Citation

Overview

PandaDock is a molecular docking suite combining a flexible-ligand conformational search with an SE(3)-equivariant Graph Neural Network scoring function that reaches Pearson R = 0.88 against experimental binding affinities on PDBbind.

The search treats every rotatable bond as an explicit degree of freedom. Position is sampled uniformly over the docking box, orientation uniformly over SO(3), and each Monte Carlo step is relaxed to a local minimum with a quasi-Newton optimizer using fully analytic gradients. Scoring runs against precomputed affinity grids, so a full search costs seconds to minutes rather than hours.

Key Features

• Flexible-ligand search — all rotatable bonds searched, uniform SO(3) orientation sampling, L-BFGS relaxation with analytic gradients
• Grid-accelerated scoring — AutoDock Vina functional form, precomputed per atom type
• PandaDock-GNN — SE(3)-equivariant scoring, R = 0.88 on PDBbind
• Hybrid workflow — search with the empirical function, rank with the GNN
• Universal rescorer — rescore poses from any docking tool (Vina, Glide, GOLD)
• Rich output — poses with bond orders, complexes, interaction analysis, plots, PandaMap 2D diagrams and an HTML report
• Specialized modes — induced-fit, metal coordination, tethered docking
• Reproducible — --seed fixes the search; every run records its parameters

Installation

Python 3.8+ is required. RDKit is a hard requirement and is most reliably installed from conda-forge.

conda create -n pandadock python=3.10
conda activate pandadock
conda install -c conda-forge rdkit

pip install pandadock

From source:

git clone https://github.com/pritampanda15/PandaDock.git
cd PandaDock
pip install -e .

Optional extras:

pip install -e ".[gnn]"      # PandaDock-GNN (PyTorch, PyTorch Geometric)
pip install pandamap         # 2D protein-ligand interaction diagrams

Quick Start

# 1. Prepare structures (adds hydrogens, assigns protonation)
pandadock-prepare -r receptor.pdb -l ligand.sdf -o prepared/

# 2. Define the binding site
pandadock-gridbox -r receptor.pdb -m similarity \
                  --reference-ligand known_ligand.sdf -o box.json

# 3. Dock
pandadock dock -r receptor.pdb -l ligand.sdf -g box.json -o results/

Or give coordinates directly:

pandadock dock -r receptor.pdb -l ligand.sdf \
               --center 10 12 8 --box 22 22 22 -o results/

Docking

pandadock dock

Exhaustiveness

Unless set explicitly, exhaustiveness scales with the number of rotatable bonds:

A budget that is ample for a rigid fragment leaves a flexible ligand's search space under-explored, and the failure is silent — the run returns a confident-looking pose from a local minimum well above the global one. Raise -e for large or highly rotatable ligands.

Two defaults worth knowing: runs are not reproducible without --seed, and --fast drops to exhaustiveness 2, so its poses should not be reported as results.

Number of poses

--num-poses is an upper bound. Poses are clustered at 2 Å heavy-atom RMSD, so a pocket supporting only a few distinct modes returns fewer than requested. The pose_diversity.png plot shows which case you are in.

Box size

A snug box outperforms a large one. Measured on confirmed search failures, reducing padding around the ligand from 8 Å to 5 Å lowered median best-of-N RMSD from 5.11 Å to 2.48 Å — and beat quadrupling exhaustiveness in a larger box, while running faster. Search volume matters more than sampling budget.

Output

Every run writes:

Prefer poses.sdf for downstream work: PDB cannot represent bond orders or formal charges, so viewers infer bonds from distance and routinely mis-assign aromatic rings.

Interactions are detected with explicit chemistry — donor/acceptor matching including the protein backbone, hydrophobic typing, electrostatics, π-stacking, π-cation and metal coordination — not distance thresholds alone.

The score is an empirical docking score in kcal/mol. It ranks poses. It is not a measured binding free energy and should not be converted to a Kd or IC50 and reported as a potency prediction; use PandaDock-GNN for affinity.

Hybrid Docking with GNN Rescoring

The recommended workflow: search with the empirical function, rank with the GNN.

pandadock gnn download-model                    # ~82 MB
pandadock hybrid -r receptor.pdb -l ligand.sdf \
                 --center 10 12 8 --box 22 22 22 \
                 -m models/pandadock_gnn.pt -o results/

Produces the same output set as dock, ranked by predicted pEC50 with the empirical score retained alongside for comparison.

GNN commands

Specialized Docking

# Induced-fit: refines receptor side chains around the ligand
pandadock-flex -r receptor.pdb -l ligand.sdf --center 10 12 8 --radius 12 -o results/

# Metal coordination: metal-aware scoring with geometry constraints
pandadock-metal dock -r receptor.pdb -l ligand.sdf --center 10 12 8 --box 22 22 22 -o results/

# Tethered: restrains the ligand centroid near a reference pose
pandadock-tethered dock -r receptor.pdb -l ligand.sdf \
                        --ref reference.sdf -t 3.0 -o results/

Tethered docking applies a flat-bottom centroid restraint inside the objective: free movement within the radius, harmonic cost beyond it.

Induced-fit docking redocks into every refined receptor, so cost scales with the number of poses carried into refinement: budget hours rather than minutes for a single ligand. Use --initial-poses-to-retain to control that trade-off. Output matches dock: poses.sdf, ligand poses, and complexes with the refined receptor. Refined receptors are written to a temporary directory and removed when the run ends, including on failure. The IFD score is the binding energy plus a weighted receptor-strain penalty, so a pose requiring more side-chain rearrangement scores worse than an equivalent one that requires none.

Metal parameters fall back to built-in approximations when no AutoDock-format parameter file is supplied. Those are adequate for identifying coordination geometry but not for quantitative metal binding energies; pass a parameter file for that.

Utilities

Grid box modes

For redocking or any case where a ligand pose is known, use similarity --reference-ligand. Cavity detection may centre the box on a neighbouring pocket, placing the true site near the box edge where sampling is poorer.

Algorithm Names

pandadock is the flexible-ligand search. Legacy names are retained for backwards compatibility and all resolve to the same algorithm:

Docking runs on CPU. The workload parallelises across ligands rather than within a single search, so throughput scales with core count.

Measuring Pose Accuracy

benchmarking/ contains a redocking harness. These scripts are development tooling and are not shipped in the PyPI package — clone the repository to use them.

# Split whole PDB entries into receptor/ligand pairs
python benchmarking/prepare_complexes.py --input complexes/ --output prepared/

# Check for ligand leakage before spending compute
python benchmarking/validate_prepared.py --manifest prepared/manifest.csv

# Redock and measure
python benchmarking/redock_benchmark.py --manifest prepared/manifest.csv \
       --output results/ --padding 5 --seed 42 -j 8

# Publication tables (Markdown and LaTeX)
python benchmarking/make_report.py results/redock_results.csv --output results/report

RMSD is symmetry-corrected and computed without superposition. Top-1 and best-of-N success rates are reported separately: quoting best-of-N as though it were top-1 substantially overstates accuracy.

analyze_redock.py separates search failures from ranking failures. A complex where a sub-2 Å pose was generated but not ranked first is a scoring problem that more sampling cannot fix, and it is the quantity that tells you whether a learned rescorer is worth applying.

Affinity Prediction Performance

These are PandaDock-GNN scoring results. For pose-prediction accuracy see Measuring Pose Accuracy above.

PDBbind v2020 Refined Set (5,316 complexes)

ULVSH Dataset (942 compounds, 10 protein targets)

BindingDB Dataset (8,891 protein-ligand complexes)

Note: Combined training with PDBbind shows reduced performance due to affinity scale differences (pKd vs pEC50). For best results, train on datasets with compatible affinity measurements.

Key Results:

• PandaDock-GNN achieves R = 0.88 on PDBbind (5,316 complexes)
• R = 0.81 on BindingDB test set (889 complexes)
• 5.5x improvement over the best baseline (VM2) on ULVSH
• Activity classification AUC = 0.94 on ULVSH test set

Documentation

Full documentation available at pandadock.readthedocs.io:

• Installation Guide
• GNN Overview
• Training Guide
• Hybrid Docking
• CLI Reference

Citation

If you use PandaDock in your research, please cite:

@article{panda2024pandadock,
  title={PandaDock: SE(3)-Equivariant Graph Neural Network Scoring for Molecular Docking},
  author={Panda, Pritam Kumar},
  journal={bioRxiv},
  year={2024},
  note={Manuscript in preparation}
}

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

License

PandaDock is released under the MIT License. See LICENSE for details.

Contact

Author: Pritam Kumar Panda Affiliation: Stanford University Email: pritampanda@stanford.edu GitHub: @pritampanda15

Acknowledgments

PandaDock builds upon excellent open-source projects:

• AutoDock Vina (scoring function inspiration)
• PyTorch and PyTorch Geometric (GNN framework)
• RDKit (molecular handling)
• E(n)-Equivariant GNN (Satorras et al. 2021)