Khalilbraham/sc-shortcutbench-public

SC-ShortcutBench: A Conflict-Row Benchmark for Metadata Shortcut Reliance in Single-Cell Foundation Models

Dataset Overview

SC-ShortcutBench is a comprehensive benchmark for evaluating whether single-cell foundation models rely on metadata shortcuts rather than expression signals for prediction. The benchmark contains ~900K cells from CELLxGENE Census with conflict-row annotations that expose metadata-biased decision-making.

Key Statistics

• Total cells: 899,801 (from CELLxGENE Census)
• Studies: 457 unique datasets
• Tissues: 26 anatomical categories
• Diseases: 28 clinical conditions
• Cell types: 200+ fine-grained categories
• Conflict rows: ~49.8K (balanced) + ~26K (decorrelated)
• Models evaluated: 10 foundation models (5 encoders + 5 generative)

Dataset Purpose

The benchmark identifies cases where single-cell foundation models exhibit shortcut reliance – predicting target labels by relying on metadata proxies rather than expression information. This is critical for understanding model robustness and identifying systematic biases in single-cell AI systems.

Files

Expression Data

• balanced_expression.h5ad (340 MB)

  • 19,250 cells with balanced split properties
  • Preserves training-time marginal distribution P(Z)
  • Safe for model training and analysis

• decorrelated_expression.h5ad (268 MB)

  • 14,420 cells with decorrelated split properties
  • Breaks correlation between true label Y and metadata shortcut Z
  • Stronger test for shortcut reliance

Results

• downstream_reliance_table.csv
  • Model predictions on conflict rows
  • Truth accuracy, shortcut agreement, TSM (Truth-Shortcut Margin)
  • 62 rows: 5 models × 3 tasks × 2 splits + controls

Experimental Design

Tasks

1. Cell Type Prediction (200+ classes)

   • Fine-grained cellular identity classification
   • Metadata proxy: Cell type labels in source studies

2. Tissue Prediction (26 classes)

   • Anatomical/organ classification
   • Metadata proxy: Tissue origin annotations

3. Disease Prediction (28 classes)

   • Clinical disease state identification
   • Metadata proxy: Disease status labels

Conflict-Row Construction (Algorithm 1)

Conflict rows are cells where:

• Training prior: P(true_label | metadata) ≠ uniform
• Test conflict: True label ≠ predicted metadata shortcut

This creates a "forced choice" between:

• Following the learned metadata shortcut (high agreement with prior)
• Relying on expression signals (low agreement with prior)

Splits

Balanced Split:

• Preserves P(Z) from training
• Conflict rows retain metadata distribution
• Tests shortcut reliance with realistic priors

Decorrelated Split:

• Y uniform within each Z stratum
• Maximizes shortcut-truth divergence
• Tests strongest shortcut preference

Evaluation Metrics

Truth-Shortcut Margin (TSM)

TSM = Accuracy(true label) - Accuracy(metadata shortcut)

• TSM < 0: Model prefers shortcut over expression
• TSM > 0: Model relies on expression over shortcut
• TSM ≈ 0: No preference (chance-level)

Key Results

All evaluated models show consistent shortcut preference:

• Truth accuracy: 15-25% (mostly random guessing)
• Shortcut agreement: 50-55% (above random)
• TSM: -24 to -40 percentage points (negative across all models)

This indicates universal shortcut reliance in current single-cell foundation models.

Models Evaluated

Encoder Models (5)

1. scFoundation - 3072-dim embeddings
2. Geneformer - 768-dim embeddings (v2, 104M params)
3. UCE - 1280-dim embeddings (4-layer)
4. scGPT - 512-dim embeddings (human-trained)
5. scPoli - 64-dim embeddings

Generative Models (5)

1. Cell2Sentence - 410M parameter model
2. Cell2Text - LLaMA-3.2 1B base
3. CellWhisperer - CLIP-based retrieval
4. scGPT - Generative mode

Data Format

AnnData Objects

• X: Gene expression matrix (log-normalized counts)
• var: Gene metadata (ENSEMBL IDs, gene names)
• obs: Cell-level metadata
  • cell_type: Fine-grained cell type
  • tissue_general: Anatomical tissue
  • disease: Disease status
  • dataset_id: Source study identifier
  • tissue_ontology_term_id: UBERON ontology
  • cell_type_ontology_term_id: CL ontology
  • Additional: donor_id, sex, assay, development_stage, etc.

CSV Results

• Columns: model, task, split, n_cells, truth_accuracy, shortcut_accuracy, TSM, CI_lower, CI_upper
• Statistical: 95% grouped-bootstrap confidence intervals (grouped by dataset_id)

Usage

Load Expression Data

import scanpy as sc

# Load balanced split
adata_balanced = sc.read_h5ad("balanced_expression.h5ad")
print(f"Cells: {adata_balanced.n_obs}, Genes: {adata_balanced.n_vars}")

# Access metadata
print(adata_balanced.obs.columns)

Load Results

import pandas as pd

results = pd.read_csv("downstream_reliance_table.csv")

# View TSM values for all models
print(results[['model', 'task', 'split', 'TSM']])

# Filter for decorrelated tissue
tissue_decorrelated = results[
    (results['task'] == 'tissue_general_prediction') &
    (results['split'] == 'decorrelated')
]
print(tissue_decorrelated)

Train and Evaluate

from sklearn.linear_model import LogisticRegression
import numpy as np

# Load data
adata = sc.read_h5ad("balanced_expression.h5ad")

# Train on all cells
X = adata.X  # Expression matrix
y = adata.obs['tissue_general'].values

model = LogisticRegression(max_iter=1000)
model.fit(X, y)

# Evaluate predictions
accuracy = model.score(X, y)

Metadata (Croissant)

See croissant.json for complete Croissant metadata specification.

License

CC-BY-4.0 with additional terms:

You are free to:

• Share and use the dataset
• Adapt and modify
• Use for commercial purposes

You must:

• Give credit to authors
• Acknowledge CELLxGENE Census as data source
• Cite the paper when using in research
• List any modifications made

Additional: Raw expression data governed by CELLxGENE Terms of Service. See individual study metadata for per-dataset licenses.

Citation

@article{sc_shortcutbench_2026,
  title={SC-ShortcutBench: A Conflict-Row Benchmark for Metadata Shortcut Reliance in Single-Cell Foundation Models},
  author={Your Author Names},
  journal={Proceedings of the Conference on Neural Information Processing Systems},
  year={2026},
  volume={39},
  pages={...},
  note={Datasets and Benchmarks Track}
}

Data Source & Acknowledgments

CELLxGENE Census

• Source: CZI Single Cell CensusData
• Access: https://cellxgene.cziscience.com/
• Reference: CZI Collective Biology Program

All cells are derived from publicly available studies in the CELLxGENE Census. See dataset metadata for per-study attributions and publications.

Funding

[Include your funding sources]

Questions & Support

For questions about:

• Dataset: [Your contact email]
• CELLxGENE Census: https://www.cziscience.com/
• Issues/Feedback: [GitHub issues link]

Dataset Card Metadata

Field	Value
Name	SC-ShortcutBench
Version	1.0
Created	2026-05-05
License	CC-BY-4.0
Language	en
Task	Single-cell analysis, bias detection
Modality	Transcriptomics (scRNA-seq)
Organism	Human
Size	~900K cells, 61K genes
Benchmarks	10 single-cell foundation models
Repository	https://huggingface.co/datasets/Khalilbraham/sc-shortcutbench-public