AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-23
Benchmarking Recent Single-Cell Synthetic Data Generation Methods on the OneK1K Dataset
The paper addresses privacy-aware sharing of donor-linked single-cell RNA-seq data. It benchmarks three synthetic data generators, scGFT, scLDM and Cell2Sentence, on the OneK1K dataset for the CAMDA 2026 ELSA challenge. Copying non-highly-variable gene values leaks heavily, whereas Poisson-based reconstruction sharply reduces membership inference risk, and scGFT and scLDM preserve biological structure better than a Poisson baseline.