unimol_docking_v2

Uni-Mol Docking V2

Uni-Mol Docking V2 Bohrium App

We update Uni-Mol Docking to Uni-Mol Docking V2, which demonstrates a remarkable improvement in performance, accurately predicting the binding poses of 77+% of ligands in the PoseBusters benchmark with an RMSD value of less than 2.0 Å, and 75+% passing all quality checks. This represents a significant increase from the 62% achieved by the previous Uni-Mol Docking model. Notably, our Uni-Mol Docking approach generates chemically accurate predictions, circumventing issues such as chirality inversions and steric clashes that have plagued previous ML models.

Service of Uni-Mol Docking V2 is avaiable at https://bohrium.dp.tech/apps/unimoldockingv2

Dependencies

• Uni-Core, check its Installation Documentation.
• rdkit==2022.9.3, install via pip install rdkit-pypi==2022.9.3 -i https://pypi.tuna.tsinghua.edu.cn/simple/ --trusted-host pypi.tuna.tsinghua.edu.cn
• biopandas==0.4.1, install via pip install biopandas tqdm scikit-learn

Data

Data	File Size	Update Date	Download Link
Raw training data	4.95GB	May 14 2024	https://zenodo.org/records/11191555
Posebusters and Astex	8.2MB	Nov 16 2023	https://github.com/deepmodeling/Uni-Mol/files/13352676/eval_sets.zip

Note that we use the Posebusters V1 (428 datapoints, released in August 2023). For the latest version, please refer to Posebusters repo.

Model weights

Model	File Size	Update Date	Download Link
unimol docking v2	464MB	May 17 2024	https://www.dropbox.com/scl/fi/sfhrtx1tjprce18wbvmdr/unimol_docking_v2_240517.pt?rlkey=5zg7bh150kcinalrqdhzmyyoo&st=n6j0nt6c&dl=0

Results

< 2.0 Å RMSD(% )	PoseBusters (N=428)	Astex (N=85)
DeepDock	17.8	34.12
DiffDock	37.9	71.76
UMol	45	-
Vina	52.3	57.65
Uni-Mol Docking	62.4	82.35
AlphaFold latest	73.6	-
Uni-Mol Docking V2	77.6	95.29

To reproduce the Posebusters results, we provide a notebook interface/posebuster_demo that includes the pipeline from data processing, model inference to metric calculation.

Training

In the training script, data_path, save_dir, finetune_mol_model, and finetune_pocket_model need to be specified.

The pretrained molecular and pocket model weights can be obtained from Uni-Mol repo. We use the no_h version weights for molecule.

bash train.sh

Inference

We add an interface for model inference in interface/demo.py.

About inputs and outpus:

• --input-protein: PDB file, abusolute path or raletive path, in batch_one2one mode, list of paths

• --input-ligand: SDF file, abusolute path or raletive path; in batch mode, list of paths

• --input-docking-grid: JSON file, include center coordinate and box size, abusolute path or raletive path; in batch mode, list of paths

• --output-ligand-name: str, the output SDF file name; in batch mode, list of names

• --output-ligand-dir: str, abusolute path or raletive path

In batch mode, you can save input_protein, input_ligand, input_docking_grid, and output_ligand_name to a CSV file and use --input-batch-file to input it.

Other parameters used:

• --steric-clash-fix: The predicted SDF file will be corrected for chemical detail and clash relaxation.

• --mode: optional values are single, batch_one2one and batch_one2many.

  • single represents one protein and one ligand as input.
  • batch_one2one represents a batch of proteins and a batch of ligands, where the relationship is one-to-one.
  • batch_one2many represents one protein and a batch of ligands, where the relationship is one-to-many.

Demo:

cd interface
bash demo.sh  # demo_batch_one2one.sh for batch mode

Or refer to this notebook interface/posebuster_demo.

Citation

Please kindly cite this paper if you use the data/code/model.

@article{alcaide2024uni,
  title={Uni-Mol Docking V2: Towards Realistic and Accurate Binding Pose Prediction},
  author={Alcaide, Eric and Gao, Zhifeng and Ke, Guolin and Li, Yaqi and Zhang, Linfeng and Zheng, Hang and Zhou, Gengmo},
  journal={arXiv preprint arXiv:2405.11769},
  year={2024}
}

License

This project is licensed under the terms of the MIT license. See LICENSE for additional details.