PhysDock: A Physics-Guided All-Atom Diffusion Model for Accurate Protein-Ligand Complex Prediction

---

****************************************************************************
*    ██████╗ ██╗  ██╗██╗   ██╗███████╗██████╗  ██████╗  ██████╗██╗  ██╗    *
*    ██╔══██╗██║  ██║╚██╗ ██╔╝██╔════╝██╔══██╗██╔═══██╗██╔════╝██║ ██╔╝    *
*    ██████╔╝███████║ ╚████╔╝ ███████╗██║  ██║██║   ██║██║     █████╔╝     *
*    ██╔═══╝ ██╔══██║  ╚██╔╝  ╚════██║██║  ██║██║   ██║██║     ██╔═██╗     *
*    ██║     ██║  ██║   ██║   ███████║██████╔╝╚██████╔╝╚██████╗██║  ██╗    *
*    ╚═╝     ╚═╝  ╚═╝   ╚═╝   ╚══════╝╚═════╝  ╚═════╝  ╚═════╝╚═╝  ╚═╝    *
****************************************************************************

PhysDock is a physics-guided all-atom denoising diffusion model designed for predicting flexible protein-ligand complex structures.

This repo is Implemented by KexinZhang ([email protected]) & JialeYu ([email protected]).

Overview

PhysDock is a non-equivariant physics-guided all-atom denoising diffusion model designed for predicting flexible protein-ligand complex structures.

PhysDock Overview

Currently, PhysDock supports multi protein chains and only one small molecule ligand input.

Installation

• PhysDock needs to search the MSAs for the protein sequence of the receptor. Therefore, the corresponding tools need to be installed and five databases must be prepared.

# Install Bio Tools
apt install hhsuite hmmer

# Download Databases
sh scripts/download_homo_datasets.sh <target_databases> # For example, $HOME/libs

• We provide a conda .yaml file to create environment to run PhysDock

# Create ENV
conda env create -f enviroment.yaml

# ENV Activation
conda activate PhysDock

• Download model params from zenedo and move it to params dir.

# Download latest EMA Params
cd params
wget https://zenodo.org/records/15178859/files/params.pt

Inference

PhysDock provides convenient inference scripts for two scenarios: redocking and virtual screening (VS). Prior to redocking and VS, it is necessary to conduct corresponding preprocessing on a system. Subsequently, the cached files obtained from preprocessing are used for complex structure prediction and ranking.

System Preparation

In PhysDock, a system encompasses one or more receptor chains along with a ligand. The input is a .pdb file of receptor and a .sdf file of ligand, while the output consists of the raw features (.pkl.gz file) for model input, typically structural-related features. In addition to structural features, PhysDock generates MSA-related features, including msa_feature and uniprot_msa_feature, by searching all sequence databases. We utilize the MD5 encoding of the protein sequence as a query to reuse the MSA features. Moreover, the raw feature can also encompass key residues involved in interactions processed by PLIP, and specify high-importance key residues with a certain probability during inference.

Below is an example of system preparation of demo input system.

BASE=$(dirname $0)

# Generate systems pkl.gz
# The receptor pdb and ligand sdf and ligand CCD id are inputs.
# The output is a system `pkl.gz`
python $BASE/prepare_system.py \
  --receptor_pdb_path $BASE/demo/system_preparation/receptor.pdb \
  --ligand_sdf_path $BASE/demo/system_preparation/EJQ.sdf \
  --ligand_ccd_id EJQ \
  --systems_dir $BASE/demo/system_preparation/systems

# Get MSA features
python $BASE/run_homo_search.py \
  --input_fasta_path $BASE/demo/system_preparation/systems/fastas \
  --features_dir $BASE/demo/system_preparation/features \
  --bfd_database_path <target_databases>/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
  --uniclust30_database_path <target_databases>/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
  --uniref90_database_path <target_databases>/uniref90.fasta \
  --mgnify_database_path <target_databases>/mgy_clusters.fa \
  --uniprot_database_path <target_databases>/uniprot.fasta \
  --jackhmmer_binary_path /usr/bin/jackhmmer \
  --hhblits_binary_path /usr/bin/hhblits

Redocking

The redocking procedure can refer to the following script redocking_demo.sh.

# Run Demo
sh redocking_demo.sh

# redocking_demo.sh
BASE=$(dirname $0)

python $BASE/redocking.py \
  -i $BASE/examples/demo/Posebusters_subset \
  -f $BASE/examples/demo/features \
  --crop_size 256 \
  --atom_crop_size 2048 \
  --enable_physics_correction \
  --use_pocket \
  --use_key_res \
  --enable_ranking

# Get Help Info
python redocking.py -h

Virtual Screening

The VS procedure can refer to the following script screning.sh. Here, it should be noted that when conducting screening for a specified pocket, the position of the ligand in the system .pkl.gz file (from System Preparation procedure) will serve as the search area for the pocket.

# Run Demo
sh screening_demo.sh

Full Docking Pipeline

# Run full pipeline
sh docking_demo.sh

Other application scenarios

Other application scenarios, such as blind docking, cross docking or standard precision (SP) flexiable docking can be implemented by specifying different parameters of python scripts.

Available Data

The training dataset, the preprocessed benchmark dataset, and the prediction results presented in the paper can be obtained through the following scripts from zenodo.

Benchmark Data

• PhysDock acheives SOTA redocking results in three benchmarks.

• The preprocessed benchmark dataset can be download through following script.

# Download all the preprocessed benchmark dataset including Posebusters, DeepDockingDare and PhiBench.
sh scripts/download_benchmarks.sh

• The training and validation dataset can be download by the following scripts

sh scripts/download_dataset.sh

Citation

If you want to cite this work, please use the following bibtex

@article{zhang2025physdock,
  title={PhysDock: A Physics-Guided All-Atom Diffusion Model for Protein-Ligand Complex Prediction},
  author={Zhang, Kexin and Ma, Yuanyuan and Yu, Jiale and Luo, Huiting and Lin, Jinyu and Qin, Yifan and Li, Xiangcheng and Jiang, Qian and Bai, Fang and Dou, Jiayi and others},
  journal={bioRxiv},
  pages={2025--04},
  year={2025},
  publisher={Cold Spring Harbor Laboratory}
}

License

MIT License