Community benchmarking of generative models for privacy-preserving synthetic bulk RNA-seq data generation.

Expressive deep generative models for synthetic bulk RNA-seq across two cancer cohorts (~1,000 and ~5,000 patients) achieve strong predictive utility but are often more vulnerable to membership inference.

• Why it matters: The challenge is to preserve biological signals in high-dimensional omics while resisting adversarial privacy attacks, where the balance between utility and privacy is not well characterized.
• What they did: CAMDA 2025 benchmarked 11 generative methods on synthetic bulk RNA-seq using 978 landmark genes, scoring distributional fidelity, downstream utility, biological plausibility, and empirical privacy risk.
• The result: Differentially private methods improve resistance to membership inference at the cost of reduced utility, while simpler statistical approaches show competitive utility with moderate privacy risk, guiding model choice and multi-attack privacy evaluation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.