Benchmark Datasets: A Long-Tailed Class-Incremental Annotation Framework for Single-Cell Transcriptomics

This dataset collection provides a curated benchmark for evaluating long-tailed and class-incremental annotation methods in single-cell RNA sequencing (scRNA-seq). It includes five widely used human single-cell transcriptomic datasets spanning diverse tissues and biological contexts:

He: Human embryonic skeletal development (Cell Research, 2021)

Madissoon: Stability of lung, spleen, and esophagus after cold preservation (Genome Biology, 2020)

Stewart: Immune zonation in human kidney (Science, 2019)

Vento: Early maternal–fetal interface reconstruction (Nature, 2018)

Zheng68k: Large-scale profiling of PBMCs using 10x Genomics (Nature Communications, 2017)

These datasets have been preprocessed and harmonized to support evaluation in long-tailed class distribution scenarios and class-incremental learning pipelines.

They are used in our study proposing ZhangCIA (single-cell Long-Tailed Class-Incremental Annotation), a novel framework addressing the challenge of annotating under-represented cell types and incrementally emerging cell classes in scRNA-seq data. The datasets in this collection serve as the experimental foundation for training and evaluating ZhangCIA across diverse biological contexts.

Please cite the original publications when using the datasets:

Jian He et al., Cell Research, 2021.

E. Madissoon et al., Genome Biology, 2020.

Benjamin J. Stewart et al., Science, 2019.

Roser Vento-Tormo et al., Nature, 2018.

G. X. Y. Zheng et al., Nature Communications, 2017.

If you use this dataset in your research, please also cite our work introducing ZhangCIA (citation to be provided upon publication).