PandaDock-GNN Overview

PandaDock-GNN is an SE(3)-equivariant Graph Neural Network scoring function for protein-ligand binding affinity prediction.

Key Features

SE(3)-Equivariance

The model produces identical predictions regardless of the rotation or translation of the input complex. This is achieved through E(n)-equivariant graph neural network (EGNN) layers.

Heterogeneous Graph Representation

Protein and ligand atoms are represented as separate node types in a heterogeneous graph, allowing the model to learn distinct representations for each.

Multi-Task Learning

The model jointly predicts:

pEC50: Binding affinity (regression)
Activity: Binary classification (active/inactive)

Architecture

Input: Protein-Ligand Complex
       │
       ├─ Protein Atoms → Node Features (56 dims)
       ├─ Ligand Atoms  → Node Features (56 dims)
       └─ Interactions  → Edge Features (23 dims)
       │
       ├─ Node Encoders (separate for protein/ligand)
       │
       ├─ EGNN Layers × 6 (SE(3)-equivariant message passing)
       │   - Update node features
       │   - Update coordinates (equivariant)
       │
       ├─ Attention Pooling
       │   - Protein graph → embedding
       │   - Ligand graph  → embedding
       │
       └─ Prediction Heads
           ├─ Affinity → pEC50
           └─ Activity → probability

Node Features (56 dimensions)

Element type one-hot (10 dims)
SYBYL atom type one-hot (16 dims)
Partial charge (1 dim)
Hybridization one-hot (4 dims)
Aromaticity flag (1 dim)
H-bond donor/acceptor (2 dims)
Ring membership (1 dim)
Residue type one-hot (20 dims, protein only)
Backbone flag (1 dim, protein only)

Edge Features (23 dimensions)

Distance (1 dim)
Gaussian RBF distance encoding (16 dims)
Bond type one-hot (4 dims)
Interaction type flags (2 dims)

Benchmark Performance

ULVSH Dataset (942 compounds, 10 protein targets):

Metric	Value
Pearson R	0.82
Spearman ρ	0.80
RMSE	0.32
MAE	0.12

BindingDB Dataset (8,891 protein-ligand complexes):

Training Configuration	Test Pearson R	Test RMSE
BindingDB Only	0.81
BindingDB + ULVSH	0.79	0.96

PDBbind v2020 (5,316 complexes):

Metric	Value
Pearson R	0.88
Spearman ρ	0.88
RMSE	0.93 pK

PandaDock-GNN outperforms all baseline methods including: VM2, MMPBSA, MMGBSA, Gnina, Hyde, DeltaVina, GFN-FF, and PM6.

Usage

Training on ULVSH:

pandadock gnn train -d ULVSH/ -o models/ --epochs 100

Training on BindingDB:

python BindingDB_training/train_bindingdb.py \
    --bindingdb BindingDB_training/bindingdb_affinity.tsv \
    --output models/ --epochs 100

Combined Training (BindingDB + ULVSH):

python BindingDB_training/train_bindingdb.py \
    --bindingdb BindingDB_training/bindingdb_affinity.tsv \
    --ulvsh ULVSH/ --combined \
    --output models/ --epochs 100

Prediction:

pandadock gnn predict -m model.pt -p protein.mol2 -l ligand.mol2

Hybrid Docking (Recommended):

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

References

The EGNN architecture is based on:

Satorras, V. G., Hoogeboom, E., & Welling, M. (2021). E(n) Equivariant Graph Neural Networks. International Conference on Machine Learning (ICML).