TorchDock

PyTorch-based Differentiable Molecular Docking Framework

TorchDock reimplements classical empirical scoring functions (Vina, Vinardo) as fully differentiable PyTorch versions, enabling end-to-end gradient-based conformational search. Compared to traditional discrete sampling methods, TorchDock achieves up to ~100x speedup in complex flexible docking scenarios (including protein side-chain flexibility), and provides an early-stopping based virtual screening pipeline that achieves ~5x acceleration while maintaining high recall rates for large-scale compound library screening.

中文文档

---

Features

• Differentiable Vina / Vinardo scoring functions (pure PyTorch implementation)
• End-to-end gradient-based conformational search with SMAC global initialization
• Flexible docking: supports protein side-chain flexibility, ~100x speedup in high-dimensional scenarios
• Virtual screening with early stopping: XGBoost-based pre-scoring, ~5x acceleration
• Complete CLI toolchain: ligand/receptor preparation + docking + result conversion
• CPU / GPU heterogeneous computing support
• Plugin architecture for custom scoring functions

---

Installation

TorchDock requires Python ≥ 3.10 and OpenBabel.

Step 1: Create conda environment and install OpenBabel

conda create -n torchdock python=3.12 -y
conda activate torchdock
conda install -c conda-forge openbabel -y

Step 2: Install TorchDock

pip install torchdock

CPU-only Installation (optional, saves disk space)

The default installation includes CUDA runtime libraries (~2GB). If you don't need GPU support, you can install the CPU version of PyTorch first to save space:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install torchdock

Install from Source

git clone https://github.com/Med4Everyone/torchdock.git
cd torchdock
pip install -e .

Verify Installation

torchdock --help

Expected output:

TorchDock v0.1.0 — Differentiable molecular docking framework

Usage: torchdock <command> [options]

Commands:
  dock                 Run molecular docking.
  prepare_ligand       Convert SMILES or file input to PDBQT ligand.
  prepare_receptor     Prepare receptor PDBQT from a PDB file.
  define_box           Define a docking box from a ligand or manual coordinates.
  convert_result       Convert TorchDock PDBQT results to SDF and PDB.
  rmsd                 Calculate RMSD between docking poses and a reference.

---

Quick Start

# 1. Prepare receptor
torchdock prepare_receptor -i protein.pdb -o receptor.pdbqt

# 2. Prepare ligand
torchdock prepare_ligand -smi "CC(=O)O" -o ligand.pdbqt  # from SMILES
# or
torchdock prepare_ligand -i ligand.mol2 -o ligand.pdbqt   # from file (mol2/sdf/pdb/mol)

# 3. Define docking box
torchdock define_box -l ligand.pdbqt -o box.json           # auto-compute center from ligand
# or
torchdock define_box -c 10.0 20.0 30.0 -s 25 25 25 -o box.json  # manual center and size

# 4. Run docking (semi-flexible / flexible)
torchdock dock -r receptor.pdbqt -l ligand.pdbqt -b box.json -o result.pdbqt
torchdock dock -r receptor.pdbqt -l ligand.pdbqt -b box.json -o result.pdbqt -f  # flexible

# 5. Convert results
torchdock convert_result -i result.pdbqt -o ./output

See example/ directory for complete runnable examples.

---

Fast Docking

TorchDock provides an early-stopping based fast docking mode (--early_stop) for initial screening of large compound libraries:

• Truncated gradient optimization: performs limited optimization steps on candidate molecules to quickly assess binding potential
• XGBoost pre-scoring: only candidates with the best predicted scores undergo full docking optimization
• Efficiency gain: achieves ~5x overall speedup compared to full-library docking while maintaining 80% recall rate for top-scoring molecules

⚠️ Note: In fast docking mode, most candidate molecules only receive predicted scores without full conformational optimization. Use results as initial screening reference, and perform standard docking on high-scoring molecules of interest.

# Enable fast docking mode
torchdock dock -r receptor.pdbqt -l ligand.pdbqt -b box.json -o result.pdbqt --early_stop

---

CLI Reference

torchdock dock

Run molecular docking.

torchdock dock -r receptor.pdbqt -l ligand.pdbqt -b box.json -o result.pdbqt

Argument	Description	Required
-r, --protein_pdbqt_path	Receptor PDBQT file	✅
-l, --ligand_pdbqt_path	Ligand PDBQT file	✅
-o, --output_path	Output result file	✅
-b, --box_file_path	Box configuration JSON file	*
-bc, --box_center X Y Z	Box center coordinates	*
-bs, --box_size DX DY DZ	Box dimensions	*
-c, --config_file_path	Custom configuration file
-f, --flex	Enable flexible docking
--flex_residues	Flexible residues (e.g., A:123,A:125), auto-detect if not specified
-sc, --score_only	Score only, no search
-es, --early_stop	Enable early stopping
-d, --device	Compute device (cpu / cuda / cuda:0)
-nw, --num_workers	CPU worker processes (default: 4)
-v, --verbose	Verbose output

* Box definition: use -b for JSON file, or -bc + -bs for manual center and size (choose one).

torchdock prepare_ligand

Convert SMILES or molecular files to PDBQT ligand.

# From SMILES
torchdock prepare_ligand -smi "CC(=O)Oc1ccccc1C(=O)O" -o ligand.pdbqt

# From file (mol2 recommended)
torchdock prepare_ligand -i molecule.mol2 -o ligand.pdbqt

# Batch conversion
torchdock prepare_ligand -b ligands.csv -o ./output_dir

Argument	Description	Required
-smi, --smiles	SMILES string (single molecule)	*
-i, --input	Input file (mol2 recommended, also supports .pdb/.mol/.sdf)	*
-b, --batch	Batch CSV file (with ID and SMILES columns)	*
-o, --output	Output file or directory	✅
-s, --seed	Random seed (for 3D coordinate generation)
-d, --remove-h	Remove hydrogen atoms

* Choose one of three input methods.

torchdock prepare_receptor

Prepare receptor PDBQT from PDB file.

torchdock prepare_receptor -i protein.pdb -o receptor.pdbqt

Argument	Description	Required
-i, --input	Input PDB file	✅
-o, --output	Output PDBQT file	✅
-d, --remove-h	Remove hydrogen atoms
-nc, --no-clean	Skip protein cleaning

torchdock define_box

Define docking box.

# Auto-compute center from ligand
torchdock define_box -l ligand.pdbqt -o box.json

# Manual center specification
torchdock define_box -c 10.0 20.0 30.0 -s 25 25 25 -o box.json

Argument	Description	Required
-l, --ligand	Ligand file (.mol2/.sdf/.pdb/.pdbqt, auto-compute center)	*
-c, --center X Y Z	Manual box center	*
-s, --size SX SY SZ	Box dimensions (default: 20 20 20)
-o, --output	Output JSON file	✅
-v, --visualize	Generate box visualization PDB

* Choose one: -l for auto-center from ligand file, or -c for manual coordinates.

torchdock convert_result

Convert docking results to SDF and PDB formats.

torchdock convert_result -i result_remi.pdbqt -o ./output

Argument	Description	Required
-i, --input	Result PDBQT file	✅
-o, --output	Output directory	✅
-t, --top-k	Convert only top k conformers

torchdock rmsd

Calculate RMSD between docking results and reference conformation.

torchdock rmsd -p result.pdbqt -r reference.pdbqt

Argument	Description	Required
-p, --predicted	Predicted result PDBQT	✅
-r, --reference	Reference structure PDBQT	✅
-t, --top-k	Calculate only top k conformers
-q, --quiet	Quiet mode (only output RMSD values)

---

Python API

TorchDock supports docking via Python code:

from torchdock.pipeline.docking_runner import docking

# Basic docking
result = docking(
    protein_pdbqt_path="receptor.pdbqt",
    ligand_pdbqt_path="ligand.pdbqt",
    box_center=[15.0, 20.0, 25.0],
    box_size=[20.0, 20.0, 20.0],
    output_path="result.pdbqt",
)

# Returns: [torchdock_score, total_score, inter_score, intra_score, unbound_score]
print(f"TorchDock Score: {result[0]:.3f}")

Using configuration file:

result = docking(
    protein_pdbqt_path="receptor.pdbqt",
    ligand_pdbqt_path="ligand.pdbqt",
    box_file_path="box.json",
    output_path="result.pdbqt",
    config_file_path="config.yaml",
    device="cuda",  # Use GPU
)

Flexible docking:

result = docking(
    protein_pdbqt_path="receptor.pdbqt",
    ligand_pdbqt_path="ligand.pdbqt",
    box_center=[15.0, 20.0, 25.0],
    box_size=[20.0, 20.0, 20.0],
    output_path="result.pdbqt",
    flex=True,                          # Enable flexible docking
    flex_residues="A:123,A:125,B:45",   # Specify flexible residues (auto-detect if not specified)
)

---

Citation

If TorchDock is helpful to your research, please cite:

@software{torchdock,
  title={Coming Soon},
  author={Coming Soon},
  year={2026},
  url={https://github.com/Med4Everyone/torchdock}
}

---

Acknowledgments

TorchDock is jointly developed by the Alibaba Tongyi AI4S team and Professor Jian-Sheng Wu's group at China Pharmaceutical University, aiming to advance computational pharmaceutical research through open-source tools. Main contributors include Jingkun Hu, Junlong Liu, and Ji Ding.

License

Apache License 2.0