File size: 2,746 Bytes
d7a4179
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
521bb4c
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
---
license: cc-by-4.0
tags:
- ai4science
- aidd
- virtual_screening
- pocket_matching
pretty_name: ProFSADB
---
# ProFSADB

**Dataset Description**  
ProFSADB is a large-scale protein-ligand interaction pretraining dataset generated by simulating pocket-ligand complexes from high-resolution protein structures. It addresses the scarcity of experimentally determined protein-ligand complexes (e.g., PDB) by extracting **5+ million non-redundant pocket-pseudo-ligand pairs** through fragmentation and interaction modeling. Each complex mimics ligand-receptor interactions to enable robust pretraining for biomedical tasks like druggability prediction and ligand affinity estimation.  

**Dataset Structure**  
- **Samples**: Over 5 million complexes.  
- **Format**: PDB files containing one complex per file.  
  - Receptor chain (`R`): Pocket residues surrounding the pseudo-ligand.  
  - Ligand chain (`L`): Drug-like protein fragment acting as a pseudo-ligand.  
- **Stratified Sampling**: Aligned with the PDBBind (v2020) distribution to ensure biological relevance.  

**Creation Process**  
1. **Fragment Isolation**: Protein structures are segmented into fragments (pseudo-ligands), with terminal corrections to address peptide bond-breaking artifacts.  
2. **Pocket Definition**: Excludes the five nearest residues on each fragment side to focus on long-range interactions. Pockets are defined as residues with ≥1 heavy atom within **6Å** of the fragment.  
3. **Quality Control**: Complexes are filtered to retain only high-confidence interaction pairs.  

**Unique Advantages**  
- **Scale**: 50× larger than existing experimental complex datasets (e.g., PDB).  
- **Interaction Modeling**: Contrastive pretraining aligns pocket features with pretrained small-molecule representations.  
- **Diversity**: Leverages structural variety from protein databases to reduce data bias.  

**License**

CC-BY-4.0

**Citation**

If you use ProFSADB or the ProFSA method, please cite:

```
@inproceedings{gao2023self,  
  title={Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment},  
  author={Gao, Bowen* and Jia, Yinjun* and Mo, Yuanle and Ni, Yuyan and Ma, Weiying and Ma, Zhiming and Lan, Yanyan†},  
  booktitle={International Conference on Learning Representations (ICLR)},  
  year={2023},  
  url={https://openreview.net/forum?id=uMAujpVi9m}  
}
```

**Links**

- **Paper**: [Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment](https://openreview.net/forum?id=uMAujpVi9m)
- **Homepage**: [Project Page](https://atomlab.yanyanlan.com/project/profsa/)


# ProFSA Model Weights

The weights of our best model pretrained using the ProFSADB data is located at `checkpoint_best.pt`.