CAMDA / ELSA Health Privacy Challenge

This repository holds the code and documentation for the Health Privacy Challenge, run within CAMDA/ISMB Conference in 2025 and 2026, and its accompanying paper. The Health Privacy Challenge is organized in the context of the European Lighthouse on Safe and Secure AI (ELSA, https://elsa-ai.eu).

Active edition: None. The 2026 challenge is concluded in July, 2026.

Jump to: Challenge tracks · Paper (Data · Reproducing) · Editions & branches · Citation

Challenge tracks

The challenge ran two tracks (full details on the edition branches):

• Track I — Bulk RNA-seq. A Blue Team 🫐 vs Red Team 🍅 scheme. Blue teams build privacy-preserving generative methods for bulk gene expression; red teams run membership-inference attacks against the baselines and blue-team solutions to test their privacy robustness.

• Track II — Single-cell RNA-seq. Privacy and utility of synthetic scRNA-seq data in a multi-sample donor setting: building generative methods, proposing suitable evaluation metrics, and revealing privacy risks.

Paper

Towards Useful and Private Synthetic Omics: Community Benchmarking of Generative Models for Transcriptomics Data (biorXiv, 2026)

The paper focuses on Track I, Blue Team methods — the privacy-preserving generative models — evaluated against membership-inference attacks provided in this repo. Track II and the Red-team tasks are documented on the challenge edition branches but are outside this paper's scope.

The 2025 Blue Team submissions are the generative methods under analysis in addition to the baseline methods provided in the repo.

Participant methods included (code archived by the authors; cite their DOIs):

Method	Authors	DOI
DP-PGM	Pentyala et al. 2026	zenodo.23068244
Embedded-Diffusion	Kreuer. 2026	zenodo.22693680
NMF / P-NMF	Wicks. 2026	zenodo.22818979

Follow their instructions for model training and generation.

The paper also introduces additional metrics beyond the challenge scoring. Reproduce the baseline generative methods and evaluation with v1.0-paper (Track I, Blue team).

Datasets

• Preprocessed data: available at zenodo.22996320.
• Also available at ELSA Benchmarks Platform after registration.
• All benchmark materials are collected in the Zenodo challenge community.

We re-distribute pre-processed versions of two open-access bulk TCGA RNA-seq datasets, available through the GDC portal; use follows the NIH Data Access Policy and TCGA citation guidelines.

• TCGA-BRCA. suitable for cancer subtype prediction (5 subtypes); 1089 x 978 (individuals x landmark genes).

• TCGA-COMBINED. suitable for cancer type prediction (10 tissues / 12 projects); 4323 x 978 (individuals x landmark genes).

Reproducing the paper

Figures only. The paper's results are frozen in paper/results/ and the paper_figures.ipynb regenerates tables and figures from them:

#  install and activate the environment
micromamba create --file environment.yaml && micromamba activate health-privacy-env
#  then open paper/paper_figures.ipynb and run all cells

Regenerate from the scratch (cluster version). Regenerating the results from scratch: prepare the data, run the generative methods and evaluate (including baseline MIA). This assumes the code runs on SLURM cluster (Use GPU for training if available).

See submit_jobs/README.md for the step-by-step commands and the adjustments you'll need to make for your own cluster.

Editions

This code-base evolved with the challenges and its accompaniying paper. Each edition is frozen as a release and mirrored as a branch for browsing.

Edition	Release (frozen)	Branch
2025	challenge-2025	camda-2025
2026	challenge-2026	camda-2026
Paper	v1.0-paper	main

Citation

If you use this benchmark, please cite:

@article {{\"O}zt{\"u}rk2026,
	author = {{\"O}zt{\"u}rk, Hakime and Afonja, Tejumade and J{\"a}lk{\"o}, Joonas and Binkyte, Ruta and Rodriguez-Mier, Pablo and Lobentanzer, Sebastian and Wicks, Andrew and Kreuer, Jules and Ouaari, Sofiane and Pfeifer, Nico and Menzies, Shane and Pentyala, Sikha and Filienko, Daniil and Golob, Steven and McKeever, Patrick and Banerjee, Jineta and Foschini, Luca and De Cock, Martine and Saez-Rodriguez, Julio and Fritz, Mario and Stegle, Oliver and Honkela, Antti},
	title = {Towards Useful and Private Synthetic Omics: Community Benchmarking of Generative Models for Transcriptomics Data},
	year = {2026},
	doi = {10.64898/2026.03.02.707794},
	journal = {bioRxiv}
}

License

Code is licensed under BSD-3-Clause (see LICENSE). Data terms are stated on the Zenodo deposit (derived from public TCGA gene expression; see Data).