Evidence in cross-tenant-data-bridging pulls in two directions on IBD bridging. One strand reports a strong IBD association that its own project labels hypothesis-generating rather than clinical validation, alongside a null cross-cohort clustering score and coverage limited to tested strains 1. A second strand reports that pooled diagnosis in the ecotype analyses was structurally unidentifiable, and that selection-on-outcome leakage reduced the count of independent Tier-A candidates 2. The distinction matters because it determines whether cross-tenant IBD bridges face a maturity problem — more cohorts, more assays — or a design problem in which the contrast is not identified and must be rebuilt.
Evidence Sides
Strong association, explicitly self-limited to hypothesis generation 1 The IBD canonical correlation analysis (CCA — a method that finds the linear combinations of two variable blocks with maximal correlation) returned r=0.964, and PhageFoundry coverage was reported alongside it; the project labels both as hypothesis-generating rather than clinical validation. The same project supplies its own limits: HMP2/FRANZOSA clustering had cross-cohort LOSO (leave-one-study-out) ARI (adjusted Rand index, a cluster-agreement statistic scored against chance) = 0.000, and the cocktail covered only tested strains. On this reading the shortfall is one of scope and validation stage. 1
The pooled contrast is not identifiable to begin with 2 In the ecotype analyses, pooled diagnosis was structurally unidentifiable: 45 sub-studies had at least 10 HC (healthy control) samples, 5 had at least 10 CD (Crohn’s disease) samples, and 0 had at least 10 of both. Selection-on-outcome leakage — choosing candidates using the outcome they are then tested against — reduced independent Tier-A candidates from 33 to 3. On this reading the defect is in the design, not the validation stage. 2
Possible Reconciliations
- Hypothesis: different estimands. r=0.964 and ARI=0.000 may measure distinct quantities — within-analysis association versus cross-cohort cluster transfer — so both can hold simultaneously, and the dispute is over which statistic should carry the claim. 1
- Hypothesis: composition drives both numbers. The 45 / 5 / 0 sub-study imbalance 2 may mean any pooled axis recovered by CCA tracks study identity rather than diagnosis, which would also explain the zero cross-cohort LOSO ARI. 1
- Hypothesis: leakage is a matter of degree. The 33 → 3 reduction may indicate that a small transferable core survives outcome-blind selection, making the two positions endpoints of one severity scale rather than opposed verdicts. 2
Resolving Work
- Ingest a cohort that populates both arms at the stated threshold of at least 10 CD and at least 10 HC samples 2, then recompute cross-cohort LOSO ARI: does cluster agreement rise above 0.000? 1
- Re-fit CCA within individual sub-studies, holding study identity constant: is r=0.964 preserved under that restriction? 1
- Audit every sub-study for simultaneous ≥10 CD and ≥10 HC counts — currently 0 — to establish whether the identifiability precondition can be met at all in existing data. 2
- Re-derive Tier-A candidates under outcome-blind selection and test whether the 3 independent survivors replicate out-of-cohort. 2
- Assay the cocktail against strains outside the tested set: does measured coverage extend beyond the tested strains? 1