Sub-cellular localisation prediction dataset: Dealing with leakage and generalisation

Ouso, Daniel 1 2 ; Pollastri, Gianluca1

Published Jun 16, 2026 on Dryad. https://doi.org/10.5061/dryad.2ngf1vj1t

Data files

Jun 16, 2026 version files 6.18 MB

README.md

4.75 KB
scldata.tar.gz

6.17 MB

Abstract

This dataset supports research on protein subcellular localisation (SCL) prediction using deep learning (DL). It is derived from the latest release of the Universal Protein Knowledgebase (UniProtKB) as of 2023-01-24, and has undergone extensive preprocessing to improve data quality, diversity, and structure. The dataset is provided in tabular format consisting of three columns: a persistent UniProtKB identifier ("entry"; string), the protein sequence ("seq"; string), and the associated subcellular location ("scl"; string).

Two dataset streams are included: one split into training, validation, and independent test sets, and another prepared for five-fold cross-validation. The independent test set (heldout) is a common split. Associated metadata are provided in JSON format: "entries.json" (entry index), "labels.json" (bi-directional label mappings), and "splits.json" (partition structure).

The dataset is designed to support the development of robust, generalisable machine learning SCL predictors, while minimising data leakage and promoting reproducibility. It is made available under the Creative Commons Zero (CC0 1.0) licence, allowing unrestricted reuse.

Also, an open-source Python data package, p-scldata, was developed to provide convenient access to the dataset. It is available on the official Python Package Index (PyPI). Upon installation, the dataset is automatically downloaded, and users can preview the data or load specific partitions programmatically. The package is under active development, with additional features planned for future releases.

Dataset DOI: 10.5061/dryad.2ngf1vj1t

Description of the data and file structure

Deep learning (DL) has advanced computational genome annotation tasks such as protein sub-cellular localisation (SCL) prediction. Nonetheless, its potential remains underutilised, primarily due to the limited availability of high-quality reference data and the suboptimal input preparation strategies. In this submission, we develop a high-quality dataset derived from the latest release of the universal protein knowledgebase (UniProtKB), designed to address existing challenges and support trustworthy and robust SCL modelling.

This dataset contributes a strong benchmark for tackling challenges in subcellular localisation prediction modelling, facilitating improved standards in protein localisation research.

Files and variables

File: scldata.tar.gz

Interface: Python package and CLI

Description: Subcellular localisation data (scldata)

Structure: unarchived and decompressed

scldata
└── data
    ├── entries.json # contains the UniProtKB unique "entry" identifier index
    ├── labels.json # contains the mined UniProtKB  "cellular component" location bi-directional index: 13 locations; 9 single and 4 multi
    ├── scl2205.csv # contains the full dataset in comma-separated-values (.csv) format
    └── splits.json # contains the partitioning of the full dataset as index lists of records constituting each partition

Components:

entries.json: key=value pairs; key is integer index, value is UniProtKB unique identifier.
labels.json: nested key=value pairs; key is integer index, value is subcellular location for the index_to_label key-index, and vice versa for the label_to_index key-index. A binary multi-label label_to_index_multi is also provided.
scl2205.csv:
- Column 0: Table index; name, entry; type string
- Column 1: Protein sequence (X variable); name, seq; type string
- Column 2: Subcellular location (Y variable); name scl; type string
splits.json
- Abbreviations
  - cv: cross-validation (the 5-fold cross-validation dataset)
  - fk: fold k, k ranges from zero to four
  - trn: training split
  - tst:the heldout evaluation set
  - vld: validation split (for monitoring training) aka development or discovery
  - [<int index>, ...]: list of integer indices representing partition members
Classes:
1. Cytoplasm (CYT)
2. Plastid (PLA)
3. Secreted (SEC)
4. Mitochondrion (MIT)
5. Membrane (MEM)
6. Peroxisome (PER)
7. Nucleus (NUC)
8. Cell projection (CEP)
9. ER (ER)
10. Cytoplasm;Nucleus (CYT;NUC)
11. Centrosome;Cytoplasm;Cytoskeleton;Microtubule organizing center (CEN;CYT;CYTS;MTOC)
12. Cytoplasm;Membrane (CYT;MEM)
13. Cytoplasm;Cytoskeleton (CYT;CYTS)

splits
├── cv
│   ├── f0
│   │   ├── trn = [<int index>, ...] # counts: 15187
│   │   └── vld = [<int index>, ...] # counts: 1256
│   ├── f1
│   │   ├── trn = [<int index>, ...] # counts: 15203
│   │   └── vld = [<int index>, ...] # counts: 1240
│   ├── f2
│   │   ├── trn = [<int index>, ...] # counts: 15185
│   │   └── vld = [<int index>, ...] # counts: 1258
│   ├── f3
│   │   ├── trn = [<int index>, ...] # counts: 15210
│   │   └── vld = [<int index>, ...] # counts: 1233
│   └── f4
│       ├── trn = [<int index>, ...] # counts: 15265
│       └── vld = [<int index>, ...] # counts: 1178
├── vld = [<int index>, ...] # counts: 15183
├── trn = [<int index>, ...] # counts: 1260
└── tst = [<int index>, ...] # counts: 2631

Notes:

No missing values

Code/software

An open-source Python data package, p-scldata, was developed to provide convenient access to the dataset. It is available on the official Python Package Index (PyPI). Upon installation, the dataset is automatically downloaded, and users can interact with it through the command line or as a Python package (scldata). The package is under active development, with additional features planned for future releases.

Access information

Other publicly accessible locations of the data:

PyPI: https://pypi.org/project/p-scldata/

Data was derived from the following sources:

UniProtKB (release 2022_05; 20230124): https://www.uniprot.org/