OSAConsensus
OSAConsensus curates every public human OSA transcriptomic series and animal CIH model series we could verify (34 GEO series, 1,032 samples, plus one external single-cell adenoid atlas), reanalyses them under one pipeline, and asks three questions with pre-registered discipline: is there a reproducible human peripheral program, does any animal CIH model reproduce the human profile, and do the answers survive leave-one-cohort-out, permutation and held-out validation? Several headline answers are negative — and they are reported as results, with the receipts on this site.
Resource at a glance
Step 7 grand adjudication of the 22 headline conclusions (Steps 3–6)
19 of 22 headline conclusions stand after leave-one-cohort-out, 10,000× permutation, old-platform sensitivity and held-out validation (3 of them strengthened). Two were downgraded (the P283 asset claim; the rat-lung day-7 phase claim) and one was revoked (the adipose cross-cohort consensus). The full table is on the Methods & adjudication page.
Human blood: a significant discordance, not a weak signal
Only two adult blood cohorts carry an independent OSA-vs-control contrast (GSE75097, GSE283703). Across 19,135 shared genes their logFC profiles are significantly anti-correlated (Spearman ρ = −0.070, gene-label permutation p < 1e-4; nominal-gene concordance 44.9%, significantly below chance). No FDR-level cross-cohort blood consensus program exists; the tiered sets are Tier A = 0, Tier B = 1 gene, Tier C = 17 genes.
P283: internally consistent, cross-cohort irreproducible
P283 (121 genes) is an FDR-level program of the single cohort GSE283703. Inside that cohort it is coherent (CPAP long-term reversal, CD14-monocyte localisation). In the pre-declared held-out cohort GSE293433 it failed and reversed: direction concordance 20/114 = 0.175, program score higher in healthy controls (Cohen's d = −1.22, label-permutation p = 0.988). It is labelled accordingly everywhere on this site and must not be cited as a validated OSA signature.
Adipose layer: cross-cohort consensus revoked (Step 7, item F-06)
The former 1,408-gene adipose Tier B set has FDR support in 0 of 1,408 genes in the second adipose cohort (GSE38792); it rests entirely on GSE135917, whose disease contrast is a global expression shift (47.3% of genes FDR-significant, median logFC −0.134). It is presented here only as a GSE135917 single-cohort set with a permanent global-shift warning; its WGCNA modules are single-cohort exploratory observations.
Animal layer: one true consensus, one honest empty set
Mouse hippocampus yields a strict two-cohort consensus of 100 genes (Tier A 14 + B 86; the genome-wide logFC correlation is ≈ 0, so the consensus rests on double-significance intersection — both facts are shown together). Mouse liver, across four cohorts, has an empty strict consensus: the signals are real but protocol/diet-specific; a 46-gene ≥3/4 sensitivity set is provided and labelled as such. At chronic endpoints the Hallmark hypoxia ssGSEA score is paradoxically down, not up (score-level statement; gene-mean logFC ≈ 0).
Cross-species bridge: pathway fidelity is negative; acute–chronic decoupling is cross-species
Animal chronic CIH pathway profiles do not reproduce the human chronic OSA blood profile (per-cohort profile correlations median r = −0.43; zero of 13 cohorts agree on all four program signs; no combination reaches permutation significance — with only 4 programs the permutation space is 24, so the ranking is positional only and no model may be called “most human-like”). The human reference profile is itself weak (no program at FDR < 0.05 in either backbone cohort) and is flagged wherever the ranking is shown. At the gene layer, the strongest true transfer is P283 × mouse lung GSE21409 (concordance 0.557, FDR 0.023, small magnitude). Acute-vs-chronic hypoxia decoupling holds in both species.
Rat lung day-7 phase claim: registry-level description
The claim that the rat-lung response concentrates at day 7 rests entirely on a 2008 microarray platform (GSE8705). Step 7 downgraded it to a registry-level description with the platform year attached.