SynOmicsBench — Automatic Benchmark Report

Synthetic data generation, ranked across fidelity, biological utility and privacy.

🏆 Most balanced method

Gaussian Copula

Total rank-derived score 5.61 — lower is better

Methods

Seeds

5 (0, 1, 2, 3, 42)

Dimensions

Candidates ranked

30 (6×5)

Dimensions: Univariate, Bivariate, DGE, GSEA, ssGSEA, Cell Deconvolution, Survival Analysis, Privacy.
Multi-seed run — 30 candidates ranked per dimension; each method's score is the mean rank of its own seeds. DGE, GSEA, Privacy collapsed to a per-method mean before ranking (manuscript-compatible).

1 · Per dimension

How each method performs on every individual metric.

Per-metric scores (higher = better)
Method	Univariate	Bivariate	DGE	GSEA	ssGSEA	Cell Deconvolution	Survival Analysis	Privacy
Avatars K5	0.857	0.934	0.588	0.395	0.681	0.547	0.920	0.848
Avatars K10	0.843	0.932	0.579	0.432	0.633	0.522	0.925	0.853
CTGAN	0.767	0.916	0.475	0.154	0.718	0.276	0.774	0.848
Gaussian Copula	0.893	0.938	0.628	0.278	0.862	0.461	0.941	0.858
Synthpop	0.924	0.925	0.530	0.308	0.724	0.650	0.781	0.611
TVAE	0.802	0.908	0.564	0.514	0.649	0.361	0.849	0.839

2 · Per pillar

The eight dimensions roll up into three equally-weighted pillars: Broad Utility (fidelity), Narrow Utility (biology), and Privacy.

Pillar sub-scores & total (lower rank-derived score = better)
Method	Broad Utility	Narrow Utility	Privacy	Total
Gaussian Copula	1.833	3.440	0.333	5.607
Avatars K10	5.033	4.227	2.000	11.260
Avatars K5	3.633	3.680	5.333	12.647
Synthpop	3.500	5.040	8.667	17.207
CTGAN	8.433	7.693	3.667	19.793
TVAE	8.567	5.307	7.000	20.873

3 · Overall meta-rank

Pillars combined into one balanced score. Lower = more balanced across utility and privacy.

Overall ranking (lower Total = more balanced)
#	Method	Total
1 🥇	Gaussian Copula	5.607
2 🥈	Avatars K10	11.260
3 🥉	Avatars K5	12.647
4	Synthpop	17.207
5	CTGAN	19.793
6	TVAE	20.873