Datasets:

jiangteam
/

DecoyDB

	---
	task_categories:
	- feature-extraction
	language:
	- en
	tags:
	- biology
	- chemistry
	- drug
	license: apache-2.0
	size_categories:
	- 1M<n<10M
	---
	🔧[Code](https://github.com/spatialdatasciencegroup/DecoyDB), 📂[Dataset](https://huggingface.co/datasets/jiangteam/DecoyDB)
	## Dataset Summary
	DecoyDB is a curated dataset of high-resolution protein-ligand complexes and their associated decoy structures. It is designed to support research on graph contrastive learning, binding affinity prediction, and structure-based drug discovery. The dataset is derived from experimentally resolved complexes and refined to ensure data quality.

	Predicting the binding affinity of protein-ligand complexes plays a vital role in drug discovery. Unfortunately, progress has been hindered by the lack of large-scale and high-quality binding affinity labels. The widely used PDBbind dataset has fewer than 20K labeled complexes. Self-supervised learning, especially graph contrastive learning (GCL), provides a unique opportunity to break the barrier by pretraining graph neural network models based on vast unlabeled complexes and fine-tuning the models on much fewer labeled complexes. However, the problem faces unique challenges, including a lack of a comprehensive unlabeled dataset with well-defined positive/negative complex pairs and the need to design GCL algorithms that incorporate the unique characteristics of such data. To fill the gap, we propose DecoyDB, a large-scale, structure-aware dataset specifically designed for self-supervised GCL on protein–ligand complexes. DecoyDB consists of high-resolution ground truth complexes and diverse decoy structures with computationally generated binding poses that range from realistic to suboptimal. Each decoy is annotated with a Root Mean Square Deviation (RMSD) from the native pose. We further design a customized GCL framework to pretrain graph neural networks based on DecoyDB and fine-tune the models with labels from PDBbind. Extensive experiments confirm that models pretrained with DecoyDB achieve superior accuracy, sample efficiency, and generalizability.



	## Data Structure
	Each protein-ligand complex is stored in a nested directory under DecoyDB/, using the format:
	```plaintext
	DecoyDB
	├── README.md # This file
	├── complexes.csv # Complex ID and paths for the data
	├── data.zip # Structures for proteins, ligands and decoys
	├── {prefix}/ # {prefix} = first 2 characters of the complex ID (e.g., '1A', '2B')
	│ └── {complex_id}/ # Unique identifier for each complex (e.g., 1A2C_H1Q)
	│ ├── {complex_id}_ligand.pdbqt # Ligand structure in AutoDock format
	│ ├── {complex_id}_target.pdbqt # Protein structure in AutoDock format
	│ ├── {complex_id}_decoys.pdbqt # Concatenated decoy structures
	│ └── {complex_id}_decoys_scores.csv # Corresponding RMSD scores for each decoy
	```
	The data.zip can be found in assets branch.
	## Dataset Details
	### Dataset Refinement
	To construct DecoyDB, we first filtered protein–ligand complexes from the Protein Data Bank (PDB) with a resolution ≤ 2.5 Å and applied the following refinement steps:

	- Removed ligands with molecular weights outside the (50, 1000) range.
	- Excluded complexes involving metal clusters, monoatomic ions, and common crystallization molecules.
	- Retained ligands with elements limited to C, N, O, H, S, P, and halogens.
	- Retained those protein chains with at least one atom within 10 Å of the ligand.
	- Saved the ligand and protein separately.

	### Decoy Generation

	For each refined protein–ligand complex, 100 decoy poses were generated using AutoDock Vina 1.2, with a 5 Å padding grid box and an exhaustiveness parameter of 8 and remove unrealistic generated structures.

	## Dataset Statistics

	- Number of protein–ligand complexes: 61,104
	- Number of decoys: 5,353,307
	- Average number of decoys per complex: 88
	- Average RMSD: 7.22 Å
	- RMSD range: [0.03, 25.56] Å

	## Contact

	- Yupu Zhang (y.zhang1@ufl.edu), Department of Computer and Information Science and Engineering, University of Florida
	- Zhe Jiang (zhe.jiang@ufl.edu), Lead PI, Department of Computer and Information Science and Engineering, University of Florida
	- Chenglong Li (lic@ufl.edu), Department of Medicinal Chemistry, University of Florida
	- Gustavo Seabra (seabra@cop.ufl.edu), Department of Medicinal Chemistry, University of Florida
	- Yanjun Li (yanjun.li@ufl.edu), Department of Medicinal Chemistry, University of Florida

	## Citation
	If you use this dataset, please cite our paper:

	```bibtex
	@inproceedings{
	zhang2025decoydb,
	title={Decoy{DB}: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction},
	author={Yupu Zhang and Zelin Xu and Tingsong Xiao and Gustavo Seabra and Yanjun Li and Chenglong Li and Zhe Jiang},
	booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
	year={2025},
	url={https://openreview.net/forum?id=lzLo5bRgQC}
	}