|
|
---
|
|
|
license: mit
|
|
|
---
|
|
|
# Description
|
|
|
Subcellular Localization prediction is a 10-class classification task to predict where a protein locates in the cell, where each input protein *x* is mapped to a label *y* ∈ {0, 1, ..., 9}.
|
|
|
|
|
|
The digital label means:
|
|
|
|
|
|
0: Nucleus
|
|
|
|
|
|
1: Cytoplasm
|
|
|
|
|
|
2: Extracellular
|
|
|
|
|
|
3: Mitochondrion
|
|
|
|
|
|
4: Cell.membrane
|
|
|
|
|
|
5: Endoplasmic.reticulum
|
|
|
|
|
|
6: Plastid
|
|
|
|
|
|
7: Golgi.apparatus
|
|
|
|
|
|
8: Lysosome/Vacuole
|
|
|
|
|
|
9: Peroxisome
|
|
|
|
|
|
# Splits
|
|
|
|
|
|
**Structure type:** AF2
|
|
|
|
|
|
The dataset is from [**DeepLoc: prediction of protein subcellular localization using deep learning**](https://academic.oup.com/bioinformatics/article/33/21/3387/3931857). We employ all proteins (proteins that lack AF2 structures are removed), and split them based on 70% structure similarity (see [ProteinShake](https://github.com/BorgwardtLab/proteinshake/tree/main)), with the number of training, validation and test set shown below:
|
|
|
|
|
|
- Train: 10414
|
|
|
- Valid: 1368
|
|
|
- Test: 1368
|
|
|
|
|
|
# Data format
|
|
|
|
|
|
We organize all data in LMDB format. The architecture of the databse is like:
|
|
|
|
|
|
**length:** The number of samples
|
|
|
|
|
|
**0:**
|
|
|
|
|
|
- **name:** The UniProt ID of the protein
|
|
|
|
|
|
- **seq:** The structure-aware sequence
|
|
|
- **plddt**: pLDDT values at all positions
|
|
|
- **label:** classification label of the sequence
|
|
|
|
|
|
**1:**
|
|
|
|
|
|
**···** |
