Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,43 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
---
|
| 4 |
+
# SingleCell-Unseen-Benchmark
|
| 5 |
+
|
| 6 |
+
## Overview
|
| 7 |
+
A large-scale unseen single-cell transcriptomic benchmark covering tumor, stem, neural, and normal cell populations for evaluating single-cell foundation models.
|
| 8 |
+
|
| 9 |
+
## Dataset Collection
|
| 10 |
+
- **Tumor cells**: 21 cancer types from GEO (2,225 samples, 1,645,662 cells), including primary tumors, metastases, and circulating tumor cells (CTCs).
|
| 11 |
+
- **Stem cells**: 5 datasets from CELLxGENE (325,092 cells, 4 stem cell types).
|
| 12 |
+
- **Neural cells**: 1 dataset from CELLxGENE (423,707 cells, 6 neural cell types).
|
| 13 |
+
- **Normal cells**: 7 datasets from CELLxGENE (1,838,991 cells, 10 normal cell types).
|
| 14 |
+
|
| 15 |
+
All datasets were mapped to HGNC symbols, and cells with <200 detected genes were removed.
|
| 16 |
+
|
| 17 |
+
## Tumor Cell Identification
|
| 18 |
+
- 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**.
|
| 19 |
+
- CELLxGENE-derived datasets used original annotations.
|
| 20 |
+
|
| 21 |
+
## Downstream Tasks
|
| 22 |
+
Designed to evaluate foundation models across multiple cell types:
|
| 23 |
+
|
| 24 |
+
| Cell type | Task | Prediction type |
|
| 25 |
+
|-----------|------|----------------|
|
| 26 |
+
| Tumor | Tumor cell identification | Binary |
|
| 27 |
+
| Tumor | Primary site tracing | Multi-class |
|
| 28 |
+
| Stem | Stem cell identification | Binary |
|
| 29 |
+
| Stem | Stem cell subtype classification | Multi-class |
|
| 30 |
+
| Neural | Neural cell identification | Binary |
|
| 31 |
+
| Neural | Neural cell subtype classification | Multi-class |
|
| 32 |
+
|
| 33 |
+
Models take **high-dimensional embeddings** as input and predict cell types using lightweight classifiers.
|
| 34 |
+
|
| 35 |
+
## Benchmark Models
|
| 36 |
+
Evaluated models: **Geneformer, scFoundation, scGPT, UCE, scLONG**
|
| 37 |
+
|
| 38 |
+
## Evaluation Metrics
|
| 39 |
+
- Binary tasks: Accuracy, Precision, Recall, F1
|
| 40 |
+
- Multi-class tasks: Accuracy, Macro-Precision, Macro-Recall, Macro-F1
|
| 41 |
+
|
| 42 |
+
## Usage
|
| 43 |
+
Datasets can be loaded via Scanpy or other single-cell analysis frameworks.
|