metadata
license: cc-by-4.0
SingleCell-Unseen-Benchmark
Overview
A large-scale unseen single-cell transcriptomic benchmark covering tumor, stem, neural, and normal cell populations for evaluating single-cell foundation models.
Dataset Collection
- Tumor cells: 21 cancer types from GEO (2,225 samples, 1,645,662 cells), including primary tumors, metastases, and circulating tumor cells (CTCs).
- Stem cells: 5 datasets from CELLxGENE (325,092 cells, 4 stem cell types).
- Neural cells: 1 dataset from CELLxGENE (423,707 cells, 6 neural cell types).
- Normal cells: 7 datasets from CELLxGENE (1,838,991 cells, 10 normal cell types).
All datasets were mapped to HGNC symbols, and cells with <200 detected genes were removed.
Tumor Cell Identification
- GEO-derived tumor cells were re-identified using a consensus workflow: lineage-level screening based on CancerSCEM 2.0 markers, followed by malignancy confirmation using inferCNV.
- CELLxGENE-derived datasets used original annotations.
Downstream Tasks
Designed to evaluate foundation models across multiple categories:
| Category | Task | Prediction type |
|---|---|---|
| Tumor | Tumor cell identification | Binary |
| Tumor | Primary site tracing | Multi-class |
| Stem | Stem cell identification | Binary |
| Stem | Stem cell subtype classification | Multi-class |
| Neural | Neural cell identification | Binary |
| Neural | Neural cell subtype classification | Multi-class |
Models take high-dimensional embeddings as input and predict cell types using lightweight classifiers.
Benchmark Models
Evaluated models: Geneformer, scFoundation, scGPT, UCE, scLONG
Evaluation Metrics
- Binary tasks: Accuracy, Precision, Recall, F1
- Multi-class tasks: Accuracy, Macro-Precision, Macro-Recall, Macro-F1
Usage
Datasets can be loaded via Scanpy or other single-cell analysis frameworks.