Size-aware cross-dataset concordance of public transcriptomic signatures

Pipeline and derived matrices for a three-domain calibration of cross-study agreement among public transcriptomic signatures.

The analysis covers 19 human datasets: peripheral blood in COVID-19 (10 datasets), peripheral blood in sepsis (5), and aortic wall in dissection or aneurysm (4). This gives 61 pairwise signature comparisons.

Headline results. Median pairwise Jaccard overlap was 0.014, 0.072 and 0.023 across the three domains. Observed overlap was 5.6, 10.2 and 6.5 times the expectation for random gene sets of the same size, with 47 of 61 pairs significant after FDR correction. Consensus across ten COVID-19 blood datasets retained two genes from a union of 4,225, and relaxing agreement to 20 per cent retained 717. Simulations with a known shared 250-gene signature hold at 250 for every k and converge after two datasets, so the failure is unstable detection rather than sample size.

The release contains a size-aware concordance measure, a permutation null, consensus analysis across datasets, ground-truth simulations in shared-signal and private-signal regimes, and an offline reproduction bundle. No expression matrices are redistributed; all 19 datasets are public and cited by accession in the paper.