pandadock 4.0.2

Molecular docking with SE(3)-equivariant GNN scoring - achieves R=0.88 on PDBbind

PandaDock - Molecular Docking with GNN Scoring

---

---

SE(3)-Equivariant GNN Scoring for Molecular Docking

Installation | Quick Start | Documentation | Benchmark | Citation

---

Overview

PandaDock v4.0 features a novel SE(3)-equivariant Graph Neural Network (GNN) scoring function that achieves state-of-the-art correlation with experimental binding affinities (R=0.88 on PDBbind, R=0.82 on ULVSH, R=0.81 on BindingDB). The hybrid docking workflow combines traditional pose generation with GNN rescoring to deliver superior accuracy.

Key Features

• PandaDock-GNN: SE(3)-equivariant scoring achieving Pearson R = 0.88 on PDBbind
• Hybrid Docking: Combined pose generation + GNN rescoring (recommended workflow)
• Universal Rescorer: Rescore poses from ANY docking tool (Vina, Glide, GOLD, etc.)
• Vina-Style Scoring: AutoDock Vina empirical weights as default scoring
• Multi-Task Learning: Joint pKd/pEC50 regression + activity classification
• Heterogeneous Graphs: Separate protein/ligand node types with interaction edges
• Specialized Modes: Flexible, metal coordination, and tethered docking

---

Benchmark Performance

PDBbind v2020 Refined Set (5,316 complexes)

Metric	Value
Pearson R	0.88
Spearman R	0.88
RMSE	0.93 pK units
MAE	0.68 pK units
Within 1.0 pK	77.5%
Within 1.5 pK	90.5%

ULVSH Dataset (942 compounds, 10 protein targets)

Method	Type	Pearson R	N
PandaDock-GNN (test)	ML Scoring	0.82	95
PandaDock-GNN (full)	ML Scoring	0.67	942
VM2	ULVSH Baseline	0.15	942
PM6	ULVSH Baseline	0.08	939
Hyde	ULVSH Baseline	0.02	942
Gnina	ULVSH Baseline	0.01	941

BindingDB Dataset (8,891 protein-ligand complexes)

Training Configuration	Test Pearson R	Test RMSE	N (train)
BindingDB Only	0.81	-	7,113
BindingDB + ULVSH	0.79	0.96	7,866
BindingDB + ULVSH + PDBbind	0.49	1.37	12,118

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

---

Installation

Prerequisites

• Python 3.8 or higher
• Conda package manager (recommended for RDKit)

Basic Installation

# Clone repository
git clone https://github.com/pritampanda15/PandaDock.git
cd PandaDock

# Create conda environment with RDKit
conda create -n pandadock python=3.10
conda activate pandadock
conda install -c conda-forge rdkit

# Install PandaDock
pip install -e .

GNN Installation (Recommended)

# Install PyTorch and PyTorch Geometric for GNN support
pip install -e ".[gnn]"

# Or manually:
pip install torch torch-geometric torch-scatter torch-sparse

For detailed installation instructions, see INSTALL.md.

---

Quick Start

Download Pre-trained Model (Recommended)

Get started immediately with the pre-trained model:

# Download the pre-trained model (~82 MB)
pandadock gnn download-model

# Model is saved to models/pandadock_gnn_v4.pt

Hybrid Docking (Recommended)

The hybrid workflow combines traditional pose generation with GNN rescoring for best accuracy:

# Using pre-trained model
pandadock hybrid -r protein.pdb -l ligand.sdf \
                 --center 10 20 30 --box 20 20 20 \
                 -m models/pandadock_gnn_v4.pt \
                 -o results/

# Or train your own model first
pandadock gnn train -d ULVSH/ -o models/ --epochs 100
pandadock hybrid -r protein.pdb -l ligand.sdf \
                 --center 10 20 30 --box 20 20 20 \
                 -m models/best_model.pt \
                 -o results/

Traditional Docking

# Simple docking with Vina-style scoring
pandadock dock -r protein.pdb -l ligand.sdf \
               --center 10 20 30 --box 20 20 20 \
               -o results/

GNN Prediction Only

# Predict binding affinity for a pre-docked complex
pandadock gnn predict -m model.pt -p protein.mol2 -l ligand.mol2

Universal Rescorer (NEW)

Rescore poses from ANY docking tool using the GNN:

# Rescore poses from AutoDock Vina
pandadock gnn rescore -m model.pt -r receptor.pdb -p vina_out.sdf -o ranked.csv

# Rescore poses from pandadock-flex
pandadock gnn rescore -m model.pt -r protein.pdb -p flex_poses.sdf --output-sdf ranked.sdf

# Rescore poses from Glide, GOLD, or any other tool
pandadock gnn rescore -m model.pt -r protein.pdb -p docked_poses.sdf

Compare Against Baselines

# Benchmark GNN against all baseline methods
pandadock gnn compare -m model.pt -d ULVSH/ -o comparison/

---

Commands

Core Commands

Command	Description
pandadock dock	Traditional docking with Vina-style scoring
pandadock hybrid	Hybrid docking with GNN rescoring (recommended)

GNN Commands

Command	Description
pandadock gnn download-model	Download pre-trained model (~82 MB)
pandadock gnn train	Train GNN model on dataset (ULVSH, PDBbind, or combined)
pandadock gnn predict	Predict binding affinity for a single complex
pandadock gnn rescore	Universal rescorer for poses from ANY docking tool
pandadock gnn benchmark	Benchmark model performance on test set
pandadock gnn compare	Compare against baseline scoring methods

Specialized Docking

Command	Description
pandadock-flex	Flexible/induced-fit docking
pandadock-metal	Metal coordination docking
pandadock-tethered	Constrained docking near reference

Utility Tools

Command	Description
pandadock-prepare	Prepare ligands (add H, generate 3D)
pandadock-gridbox	Generate grid box configurations
pandadock-report	Generate analysis reports

---

Universal GNN Rescorer

The pandadock gnn rescore command allows you to rescore docked poses from any docking software using the SE(3)-equivariant GNN:

Supported Input

• AutoDock Vina output (SDF/PDBQT converted to SDF)
• Glide poses (SDF)
• GOLD poses (SDF)
• pandadock-flex flexible docking poses
• pandadock-metal metal coordination poses
• pandadock-tethered constrained poses
• Any multi-conformer SDF file

Usage

pandadock gnn rescore -m model.pt -r receptor.pdb -p poses.sdf [OPTIONS]

Options:
  -m, --model PATH      Trained GNN model checkpoint (required)
  -r, --receptor PATH   Receptor PDB or MOL2 file (required)
  -p, --poses PATH      Multi-conformer SDF with poses (required)
  -o, --output PATH     Output CSV with ranked poses (default: rescored_poses.csv)
  --output-sdf PATH     Output SDF with GNN scores as properties
  --site-radius FLOAT   Binding site extraction radius (default: 10 A)

Example Workflow

# Step 1: Run docking with your preferred tool
vina --receptor protein.pdbqt --ligand ligand.pdbqt --out poses.sdf

# Step 2: Rescore with PandaDock-GNN
pandadock gnn rescore -m model.pt -r protein.pdb -p poses.sdf \
    -o ranked.csv --output-sdf ranked.sdf

# Output CSV columns:
# pose_name, pose_index, gnn_pKd, gnn_energy, activity_prob, predicted_active, gnn_rank

Output SDF Properties

When using --output-sdf, each molecule gets these properties:

• GNN_pKd - Predicted pKd/pKi value
• GNN_Energy - Predicted binding energy (kcal/mol)
• GNN_Activity - Activity probability (0-1)
• GNN_Rank - Rank based on GNN score (1 = best)

---

GNN Architecture

PandaDock-GNN uses an SE(3)-equivariant heterogeneous graph neural network:

Input: Protein-Ligand Complex
  |
  +-- MOL2/PDB/SDF Parser --> Atom coordinates, types, charges
  |
  +-- Graph Builder --> HeteroData graph
  |   - Protein nodes (56 features)
  |   - Ligand nodes (56 features)
  |   - Interaction edges (23 features, 5A cutoff)
  |
  +-- EGNN Layers x 6 (SE(3)-equivariant message passing)
  |   - Coordinate updates preserve symmetry
  |   - Edge attention mechanism
  |
  +-- Attention Pooling --> Graph-level representation
  |
  +-- Prediction Heads
      - pKd/pEC50 regression
      - Activity classification (sigmoid)

Node Features (56 dims):

• Atom type one-hot (10)
• SYBYL atom type (16)
• Partial charge (1)
• Hybridization (4)
• Aromaticity, H-bond donor/acceptor (4)
• Residue type (20, protein only)
• Backbone flag (1)

Edge Features (23 dims):

• Distance (1)
• Gaussian RBF expansion (16)
• Bond type one-hot (4)
• Interaction type flags (2)

---

Scoring Functions

Function	Description	Use Case
vina	AutoDock Vina empirical scoring (default)	General docking
physics_based	Lennard-Jones + electrostatics	Detailed energy analysis

---

Output Files

Dock Command

docking_output/
+-- complex1.pdb, complex2.pdb, ...   # Protein-ligand complexes
+-- pose1.pdb, pose2.pdb, ...         # Ligand poses only
+-- docking_results.json              # Complete results with energies
+-- interaction_analysis.json         # Detailed interactions
+-- binding_affinities.png            # Affinity distribution

Hybrid Command

hybrid_output/
+-- hybrid_results.csv                # Rankings with GNN + Vina scores
+-- pose_1_pec50_X.XX.pdb             # Top poses with pEC50 in filename
+-- complex_1.pdb, ...                # Protein-ligand complexes

Rescore Command

rescored_poses.csv                    # Ranked poses with GNN scores
ranked.sdf (optional)                 # SDF with GNN properties

---

Training Your Own GNN Model

PandaDock supports training on three dataset formats: ULVSH, PDBbind, and BindingDB. For detailed dataset preparation instructions, see the Dataset Preparation Guide.

Dataset Requirements

Dataset	Format	Key Files
ULVSH	Directory	vitro.tsv + protein.mol2, ligand.mol2, site.mol2 per compound
PDBbind	Directory	INDEX_refined_data.2020 + {pdb}_pocket.pdb, {pdb}_ligand.mol2
BindingDB	TSV file	TSV with complex_id, protein_file, ligand_file, pK columns

Single Dataset Training

# Train on ULVSH (942 compounds, 10 targets)
pandadock gnn train -d ULVSH/ -o models/ --epochs 100

# Train on PDBbind (5,316 complexes)
pandadock gnn train -p PDBbind/ -o models/ --epochs 100

# Train on BindingDB (custom TSV file)
pandadock gnn train -b bindingdb_affinity.tsv -o models/ --epochs 100

Combined Dataset Training (Recommended)

Combining datasets improves generalization. Use --balanced to prevent larger datasets from dominating:

# BindingDB + ULVSH (recommended for screening)
pandadock gnn train -b bindingdb.tsv -d ULVSH/ -o models/ \
    --balanced --epochs 100

# ULVSH + PDBbind (recommended for structure-based)
pandadock gnn train -d ULVSH/ -p PDBbind/ -o models/ \
    --balanced --epochs 200

# All three datasets
pandadock gnn train -d ULVSH/ -p PDBbind/ -b bindingdb.tsv -o models/ \
    --balanced --epochs 200 --batch-size 32

Training Options

Option	Default	Description
--epochs	100	Number of training epochs
--batch-size	32	Batch size (reduce if out of memory)
--hidden-dim	256	Hidden layer dimension
--num-layers	6	Number of EGNN layers
--balanced	off	Balance sampling across datasets
--patience	20	Early stopping patience

Benchmark on Test Set

pandadock gnn benchmark -m models/best_model.pt -d ULVSH/ -o results/

---

Examples

See the examples/ directory:

• examples/basic_docking/ - Simple docking workflow
• examples/flexible_docking/ - Induced-fit docking
• examples/metal_docking/ - Metalloprotein docking

---

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: [e-mail] 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)

---

Star this repository if you find it useful!

Report Bug | Request Feature