DiffBind

DiffBindFR: SE(3) Equivariant Network for Flexible Protein-Ligand Docking

A PyTorch reproduction of DiffBindFR from:

Zhu et al., "DiffBindFR: An SE(3) Equivariant Network for Flexible Protein-Ligand Docking", Chemical Science 2024.

DiffBindFR jointly samples ligand poses and binding-pocket side-chain torsion angles using an SE(3)-equivariant diffusion generative model, enabling flexible receptor docking without exhaustive conformational search.

---

Architecture Overview

Input: Apo/Holo protein pocket + ligand graph
        │
        ├─ Protein encoder  (GVP – Geometric Vector Perceptron)
        ├─ Ligand encoder   (Graph Transformer / EGNN)
        │
        └─ SE(3)-Equivariant Diffusion Model
               │
               ├─ Ligand translation  (R³ diffusion)
               ├─ Ligand rotation     (SO(3) diffusion)
               ├─ Ligand torsion      (T^d diffusion)
               └─ Side-chain χ₁ angles (T^k diffusion)
                        │
                        ▼
               MDN Confidence Model → Top-1 pose selection

---

Repository Layout

DiffBindFR/
├── README.md
├── environment.yml           # Conda environment
├── configs/
│   └── default.yaml          # Hyperparameters
├── models/
│   ├── __init__.py
│   ├── gvp_encoder.py        # GVP protein backbone encoder
│   ├── ligand_encoder.py     # Graph Transformer ligand encoder
│   ├── diffusion.py          # SE(3) diffusion model (core)
│   ├── score_network.py      # Score / denoising network
│   └── mdn_confidence.py     # MDN confidence scoring model
├── data/
│   ├── __init__.py
│   ├── dataset.py            # PDBbind / CD cross-dock dataset
│   ├── featurize.py          # Protein & ligand featurization
│   └── transforms.py         # Data augmentation & noise schedules
├── utils/
│   ├── __init__.py
│   ├── so3.py                # SO(3) / IGSO3 utilities
│   ├── torus.py              # Torus diffusion for torsion angles
│   ├── geometry.py           # RMSD, chi-angle helpers
│   └── sampling.py           # ODE / SDE solvers
├── evaluation/
│   ├── __init__.py
│   ├── metrics.py            # L-RMSD, sc-RMSD, C-Dist, PB-success
│   └── posebusters.py        # PoseBusters wrapper
├── scripts/
│   ├── train.py              # Training entry point
│   ├── inference.py          # Docking inference
│   └── build_cd_testset.py   # CD cross-dock test set construction
└── tests/
    └── test_models.py        # Unit tests

---

Installation

conda env create -f environment.yml
conda activate diffbindfr

# Optional: install PoseBusters for validity checks
pip install posebusters

---

Quick Start

Training

python scripts/train.py \
    --config configs/default.yaml \
    --data_dir /path/to/PDBbind2020 \
    --split time_split \
    --gpus 4

Inference (flexible docking)

python scripts/inference.py \
    --protein receptor.pdb \
    --ligand ligand.sdf \
    --n_poses 40 \
    --solver SDE \
    --denoising_steps 20 \
    --confidence MDN \
    --output docked_poses.sdf

Evaluation

python evaluation/metrics.py \
    --pred docked_poses.sdf \
    --ref crystal.sdf \
    --protein receptor.pdb

---

Key Hyperparameters (configs/default.yaml)

Parameter	Value	Description
n_poses	40	Poses sampled per complex
denoising_steps	20	SDE/ODE steps
solver	SDE	Diffusion solver type
pocket_radius	10.0 Å	Binding pocket definition
mdn_gaussians	10	MDN mixture components
gvp_layers	3	GVP encoder depth
gt_layers	6	Graph Transformer depth
hidden_dim	256	Node embedding dimension

---

Results (reproduced targets)

PDBbind Time-Split Test Set

Method	L-RMSD top1	sc-RMSD top1
AutoDock Vina	~37%	–
DiffDock	~38%	–
KarmaDock	~43%	–
DiffBindFR-MDN	~51%	~72%

(Success rate = fraction with RMSD < 2 Å)

CD Cross-Dock Test Set (CASF2016 Apo-Holo)

Method	SR (RMSD<2Å)	PB-SR
Glide	13.2%	13.2%
KarmaDock	10.9%	10.9%
DiffBindFR-MDN	–	56.1%

---

Citation

@article{zhu2024diffbindfr,
  title   = {DiffBindFR: An SE(3) Equivariant Network for Flexible Protein-Ligand Docking},
  author  = {Zhu, Jintao and Gu, Zhonghui and Pei, Jianfeng and Lai, Luhua},
  journal = {Chemical Science},
  year    = {2024},
  doi     = {10.1039/D3SC06803J}
}

---

Acknowledgements

This reproduction builds on:

• GVP-GNN (Jing et al., ICLR 2021)
• DiffDock (Corso et al., ICLR 2023)
• KarmaDock (Zhang et al., Nature Comp. Sci. 2023)