setHARD: protein subcellular localization

An external, single-label localization test set derived from the Light Attention setHARD dataset. This is not DeepLoc 2.0.

Data

The test split contains 483 proteins. Seven of the original 490 proteins were excluded for exceeding the 2,048-residue limit. Retained sequences range from 52 to 2,035 residues; selenocysteine (U) is preserved.

Column	Description
id	Unique sample identifier.
sequence	Amino-acid sequence, without truncation.
targets	Integer localization class.
split	Always test.

Class indices: 0 Peroxisome; 1 Extracellular; 2 Cytoplasm; 3 Golgi apparatus; 4 Cell membrane; 5 Plastid; 6 Mitochondrion; 7 Nucleus; 8 Endoplasmic reticulum; 9 Lysosome/Vacuole.

Evaluation

Use only for external testing, not training or hyperparameter selection. Report Macro-F1 over all ten classes. Verify class-index compatibility with the training dataset before evaluating a trained localization model. Classes are imbalanced, with as few as three examples in one class.

Usage

from datasets import load_dataset

dataset = load_dataset("flair-bio/sethard")

Consult the upstream source for licensing and attribution requirements.