Methods & adjudication
Pipeline summary per step, then the full Step 7 grand adjudication table (H1–H22) that governs the wording of every claim on this site, and the inheritance list passed to the manuscript.
Step 1 — Census
All GEO series for human OSA and animal CIH were enumerated and adjudicated series by series; the census froze at 34 series (human CORE 11, animal strict CORE 14, disease-background 8, acute panel 1) plus the external adenoid atlas.
Step 2 — Download & protocol mining
34/34 series downloaded and GSM-reconciled exactly (1,032 samples). CIH protocol fields (FiO₂ nadir, cycle, daily hours, duration, recovery arms) were mined from the original papers with per-field sources; unstated fields are recorded as unstated, never imputed.
Step 3 — Human peripheral layer
Uniform reanalysis of the blood and adipose cohorts: per-cohort differential expression, tiered consensus, LM22 + xCell deconvolution, CPAP paired reversal, WGCNA, and the acute-IH bridge cohort GSE71356. Adjudicated open item F-06 (GSE135917 global shift) was deferred to Step 7 and closed there.
Step 4 — Human tissue layer
The adenoid atlas was annotated under one vocabulary (238,093 cells, 17 types) with donor-level pseudobulk severity gradients, composition tests, exploratory CellChat, a design-bounded tonsil validation (GSE274855), and PBMC program localisation (GSE214865). The genioglossus comparison is descriptive-only (source-data truncation proven at the GEO deposit in Step 5).
Step 5 — Animal layer
Per-tissue multi-cohort consensus with the same tiering discipline as the human layer, a four-program ssGSEA pathway matrix (GSVA v2), background-layer offsets against internal wild-type references, the acute panel kept separate, and single-cell layers (kidney snRNA, genioglossus sc) handled descriptively where the deposits are truncated.
Step 6 — Cross-species bridge (dual layer)
Pathway layer (main): four-program profile correlation of every animal combination against the human chronic blood reference, with label permutation. Gene layer (attached, tier-labelled): a 24,618-row per-gene cross-species response table via ortholog mapping, per-unit concordance and enrichment. CPAP vs recovery arms compared asymmetrically (direction and return proportion only).
Step 7 — Robustness & held-out validation
Readouts and thresholds for the held-out validation were written down before GSE293433 was opened (validation plan timestamped). Leave-one-cohort-out for every consensus object, 10,000× gene-set permutations, old-platform (pre-2010 / EST) inclusion–exclusion sensitivity, the F-06 adipose adjudication, and the grand adjudication table H1–H22.
Grand adjudication table (Step 7): 19 standing · 2 downgraded · 1 revoked
| # | Claim | Status | Basis |
|---|---|---|---|
| H1 | No FDR-level adult peripheral-blood OSA consensus program exists (Step 3) | Standing — strengthened | Backbone cohorts significantly anti-correlated (perm p<1e-4); concordance significantly below chance; held-out validation independently concurs. |
| H2 | P283 is the most solid program asset: single-cohort FDR program with CPAP normalisation and monocyte localisation (Steps 3/4) | Downgraded | Held-out validation failed and reversed (concordance 0.175, d=−1.22). Corrected calibre: “single-cohort internally consistent, cross-cohort irreproducible”. Its within-cohort evidence stands on its own. |
| H3 | NOM18 is a nominal-tier set only (Step 3) | Standing | Held-out not supported (0.231); pooled permutation p=0.043 is nominal and single-design driven; never load-bearing. |
| H4 | No FDR cell-fraction difference in blood deconvolution (Step 3) | Standing | No check overturned it; no old-platform exposure on the human side. |
| H5 | Program scores track monocyte fraction/state (Steps 3/4) | Standing | Two steps mutually corroborate; see H12 for the statistic caveat. |
| H6 | CPAP reversal concentrates in GSE283703; it is not a general phenomenon (Step 3) | Standing — strengthened | Gene-set permutation: P283 pooled p=0.149; only the GSE283703 long-term design is significant (p=0.0135). |
| H7 | Acute IH and chronic OSA are decoupled at the gene level (Step 3 bridge) | Standing | Cross-species version strengthened separately as H20. |
| H8 | Adipose Tier B (1,408 genes) is a cross-cohort consensus (Step 3) | Revoked | 0/1,408 genes have FDR support in GSE38792; the set rests solely on the globally shifted GSE135917. Reclassified as a GSE135917 single-cohort set; its WGCNA modules downgraded with it. |
| H9 | Adenoid severity gradient is FDR-null; nominal signal points to T cells (Step 4) | Standing | GSE274855 compilation is a non-replication, consistent with a weak tissue-layer signal. |
| H10 | CellChat MIF–CD74 axis weakens in severe disease (exploratory, Step 4) | Standing | No new test in Step 7; exploratory calibre unchanged. |
| H11 | The tonsil cohort does not replicate the adenoid program (Step 4) | Standing | Compilation completed; the “vs healthy” readout is not computable by design (no healthy arm). |
| H12 | P283 localises to CD14 monocytes (Step 4) | Standing — wording tightened | Mean-shift permutation p=0.15 vs rank test p=0.0046 — both statistics must be reported together; localisation is valid only within its source-cohort context (see H2). |
| H13 | Mouse hippocampus has a true 100-gene two-cohort consensus (Step 5) | Standing | LOO formalisation confirms intersection character (Jaccard 0.07–0.10); the genome-wide ρ≈0 caveat stands alongside. |
| H14 | Mouse liver strict consensus is empty; the 46-gene set is sensitivity-tier only (Step 5) | Standing | LOO: any leave-one-out is non-empty (3–19 genes) — “empty” is a property of the full four-cohort gate; the 46-gene set's composition is cohort-sensitive (LOO survivors 3–19). |
| H15 | At chronic endpoints the hypoxia program is down, not up; only 3 pathway-grid cells pass FDR (Step 5) | Standing — wording tightened | Not old-platform dependent; but gene-mean logFC ≈ 0 (Step 7 P5) — wording is restricted to the ssGSEA score level. |
| H16 | The rat-lung response concentrates at day 7 (Step 5, GSE8705) | Downgraded | Rests entirely on a 2008 microarray platform; downgraded to a registry-level description with platform year attached. |
| H17 | Background-layer fidelity grading at the gene layer: P301S preserves/amplifies, 5XFAD abolishes (Step 5) | Standing | Permutation supports the gene-layer units; the pre-existing note that this does not reproduce at the pathway layer stands. |
| H18 | Kidney effect is proximal-tubule-led; genioglossus is descriptive-only (Step 5) | Standing | No check touched these items. |
| H19 | Pathway fidelity: no combination is significant; the ranking is positional only (Step 6) | Standing | Settled: the permutation space is 24 by design (4 programs), minimum one-sided p ≈ 0.042; animal-side LOO is stable. F-05 closed. |
| H20 | Acute–chronic decoupling is cross-species (Step 6) | Standing | Animal LOO 0/52 flips; unchanged without old platforms; load-bearing cells (Hypoxia, Oxidative stress) invariant. Circadian/Inflammation signs depend on the human reference cohort, and the heart cell has no modern chronic counterpart after old-platform exclusion — all three caveats are mandatory in the manuscript. |
| H21 | Gene-layer transfer: P283 × mouse lung significant; adenoid nominal transfer, etc. (Step 6) | Standing — strengthened | All confirmed at 10,000 permutations (p from 0.0033 to <1e-4). |
| H22 | CPAP vs animal recovery arms are an asymmetric description only (Step 6) | Standing | Design-asymmetry calibre maintained. |
How to cite
How to cite a set from this resource: (1) use the tier label as printed; (2) P283 must be described as a GSE283703 single-cohort program with its held-out reversal attached at first mention; (3) the adipose set must be named “GSE135917 single-cohort set” with the global-shift warning; (4) pathway claims are ssGSEA-score-level statements; (5) the fidelity ranking may be quoted by position only, never as significance.