Motivation

Why Generate Synthetic scRNA‑seq Data?

Single-cell RNA sequencing (scRNA-seq) has revolutionised biology by measuring gene expression in thousands of individual cells simultaneously. However, real scRNA-seq data comes with inherent limitations:

Challenge

Impact

Rare cell types

Populations like circulating tumour cells or tissue-resident stem cells appear in tiny numbers. Downstream analyses are underpowered.

Data paucity

Experiments are expensive and time-consuming. Small datasets limit the statistical power of clustering, differential expression, and machine learning models.

Patient privacy

Sharing real patient scRNA-seq data raises ethical and regulatory concerns (GDPR, HIPAA).

Benchmarking

Developing and testing new algorithms requires diverse, controlled datasets with known ground truth.

Dropout & noise

Zero-inflated expression, amplification noise, and batch effects obscure true biological signal.

Synthetic data generation addresses all five challenges simultaneously — if the synthetic data faithfully reproduces the biological properties of real data.

The GARAGE Approach

GARAGE was designed around a specific insight: standard GANs fail on scRNA-seq data because they lose rare cell populations. Mode collapse — where a GAN produces only a few “average” cell types repeatedly — is fatal for biological applications.

GARAGE solves this with a two-stage architecture:

  1. Stage 1: A Graph Attention Network (GAT) identifies the most representative cells, with explicit priority weighting for rare cell types.

  2. Stage 2: A GAN generates synthetic cells from a hybrid input that mixes random noise with GAT-selected seed cells.

The result: synthetic data that preserves both abundant and rare cell types, with distributional fidelity validated by Wasserstein distance, MMD, and clustering metrics.

Where GARAGE Fits in the Literature

Method

Approach

Rare-cell aware?

scRNA-seq optimised?

scGAN (Marouf et al.)

WGAN-GP with deep generator

No

Yes

scVAE (Lopez et al.)

\(\beta\)-VAE with NB likelihood

No

Yes

scDiffusion (Luo et al.)

DDPM on log-normalised counts

No

Yes

LSH-GAN

Random KNN subsample + GAN

Partially

No

GARAGE

GAT-seeded GAN with priority weighting

Yes

Yes

GARAGE is the first framework to explicitly couple a graph attention model with a GAN for the purpose of rare-cell preservation.

Who Should Use GARAGE?

  • Bioinformaticians wanting to augment small or imbalanced scRNA-seq datasets.

  • Methods developers needing realistic benchmark datasets with ground-truth cell type labels.

  • Clinicians exploring privacy-preserving data sharing for multi-centre studies.

  • Researchers studying rare cell populations (cancer stem cells, circulating tumour cells, tissue-resident immune cells).

Next Steps