A graph-attentive GAN for rare-cell-aware single-cell RNA-seq data generation

A central challenge in downstream single-cell RNA sequencing (scRNA-seq) analysis is the high-dimensional, small-sample (HDSS) regime, often compounded by class imbalance from rare cell types. These factors hinder robust feature (gene) selection and cell clustering and limit the realism of samples generated by existing simulators. We introduce GARAGE, a Graph-Attentive RAre-cell aware single-cell data GEneration that augments the generator’s input with a small, attention-weighted ‘leakage’ of real cells in addition to prior noise. Specifically, we build a k-nearest-neighbour cell graph and use a graph attention network (GAT) to prioritize nodes that likely represent under-sampled (rare) subpopulations; these high-attention cell embeddings are injected into the generator input to steer synthesis toward biologically plausible regions of the data manifold while respecting cell-type proportions. This attention-guided leakage accelerates training, reduces mode dropping, and yields realistic synthetic cells that preserve rare-cell structure. Across real scRNA-seq benchmarks, GARAGE improves downstream feature selection and clustering compared with state-of-the-art baselines. In summary, GARAGE directly addresses HDSS and rarity in scRNA-seq by coupling graph attention with adversarial generation to produce high-fidelity synthetic cells that enhance downstream analyses. The corresponding software is available at: https://github.com/RitwikGanguly/GARAGE.Author summary: Researchers frequently analyze single-cell transcriptomic data to unravel the mechanisms behind complex diseases. Yet, rare but important cell types are often too scarce in patient samples to study reliably. We introduce GARAGE - Graph-Attentive RAre-cell aware single-cell data GEneration - a two-step framework that explicitly couples graph attention with adversarial learning. First, a neighborhood graph with attention weights surfaces “exemplar cells”, prioritizing rare populations. Second, we apply an attention-guided “leakage”: a small, curated set of these high-score real cells is injected into the GAN’s input alongside noise. This graph-attentive prior steers the generator to the full data manifold, accelerates and stabilizes GAN convergence, and prevents mode dropping, so rare cell types are synthesized faithfully. Across real scRNA-seq benchmarks, GARAGE produces more complete, biologically coherent datasets and consistently outperforms state-of-the-art methods in downstream feature selection and clustering. By tackling the high-dimensional, small-sample regime head-on, GARAGE delivers realistic synthetic cells that preserve rare-cell structure and diversity, providing a practical tool to power robust analyses and speed discovery.