Aging and youthful function / Observational data reanalysis
Read the gene and age-support challenge
Age support changes the gingival trend
The common-span restriction removes every bulk donor over 45. These subsets are not matched-age or matched-donor validation, and expression is not restored function.
# ECM direction disagreement is broad across genes and sensitive to age support Computed 2026-10-19. This advances the youthful-function workstream by testing whether the opposing gingival structural-matrix directions are concentrated in a few genes or depend on the sampled age distribution. No new cohort, panel, age clock or intervention was introduced. **Result:** the full-cohort fibroblast/bulk disagreement spans seven of eight fixed ECM genes in age-only fits and all eight after source-specific adjustment. Removing any one gene preserves each panel direction. However, restricting the nominal common age span removes all older bulk donors and reverses that cohort's direction for all eight genes. A different, coarse shared-decade restriction retains older adults and the original opposition. This is evidence of population-dependent descriptive associations, not a resolution into a causal aging program or matched-age validation. ## Fixed inputs, genes and demographic rules The [plan](analysis_plan.json) was specified after earlier panel results and metadata review, before this batch's gene/support outcomes. It was not preregistered. The eight equally weighted genes are unchanged: `COL1A1, COL1A2, COL3A1, COL5A1, COL6A1, FN1, DCN, LUM`. All eight occur for every eligible donor in both cohorts; no annotation substitution, missing-gene imputation or outcome-based selection is used. Existing public tables supply four separate measurements: - GSE164241 fibroblast-only raw-count `log1p(CPM)`. - GSE164241 observed and common-RNA-weight **partial captured-cell** mixtures from the [previous disjoint-compartment challenge](../../checklist-mechanisms-2026-10-19/aging/REPORT.md). - GSE173078 bulk-gingiva gene values from the [separate-cohort challenge](../../checklist-followup-2026-10-19/aging/REPORT.md). These are different estimands and assay contexts. The partial mixtures omit 386 captured cells with unknown RNA and cannot be called intact-tissue bulk. The original 13 healthy inferred GSE164241 donors and 12 source-labelled healthy GSE173078 participants remain the units; cells, genes and repeated biopsies do not become extra people. Original experiment and public sources: [Williams et al., 2021](https://doi.org/10.1016/j.cell.2021.05.013), [GSE164241](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE164241), [GSE173078](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE173078). No new source data were acquired. A [metadata-only script](freeze_populations.py) freezes the following populations before gene outcomes. Exact retained **donor IDs, ages, sex and processing labels** are preserved in [frozen populations](frozen_populations.json). | Population | GSE164241 donors / retained ages | GSE173078 donors / retained ages | | --- | --- | --- | | Original healthy cohorts | 13: 21,21,22,24,28,31,32,35,37,39,51,57,61 | 12: 28,32,37,39,40,42,42,45,63,66,68,75 | | Intersection of original observed age spans, 28–61 | 9: 28,31,32,35,37,39,51,57,61 | 8: 28,32,37,39,40,42,42,45 | | Decade bins occupied by both original cohorts: 20s/30s/60s | 11: 21,21,22,24,28,31,32,35,37,39,61 | 7: 28,32,37,39,63,66,68 | The common-span rule does **not** produce matched distributions: the retained bulk sample has nobody aged 46–61. The decade rule is also coarse support, not matching: its older cell-source bin contains one 61-year donor versus bulk donors aged 63, 66 and 68; its twenties bin contains five versus one. Neither restriction establishes improved clinical validation, overlap of unmeasured covariates, or adequate older-male representation. Separate within-source models use age only, or age plus sex and processing. Processing means Cell Ranger version for GSE164241 and batch A/B for GSE173078, with no pooled batch correction. All genes in each fit use exactly the same donors/design. Failed rank or residual-degrees-of-freedom checks would remain explicit; none occurred in these specified full fits or donor omissions. ## All eight full-cohort gene effects Slopes below are relative gene `log1p(CPM)` per decade; numerical scales are not pooled across assays. Signs are descriptive and do not establish individual-gene significance or functional effects. | Gene | Fibroblast, age only | Bulk, age only | Fibroblast, adjusted | Bulk, adjusted | | --- | ---: | ---: | ---: | ---: | | COL1A1 | +0.248 | −0.265 | +0.311 | −0.318 | | COL1A2 | +0.212 | −0.048 | +0.245 | −0.234 | | COL3A1 | +0.251 | −0.271 | +0.281 | −0.306 | | COL5A1 | +0.165 | −0.039 | +0.201 | −0.208 | | COL6A1 | +0.183 | −0.066 | +0.238 | −0.195 | | FN1 | +0.085 | +0.006 | +0.077 | −0.126 | | DCN | +0.124 | −0.288 | +0.132 | −0.339 | | LUM | +0.111 | −0.191 | +0.002 | −0.164 | The exception in age-only fits is near-zero bulk FN1; adjusted fibroblast LUM is also near zero. These small signs should not be overread. Nevertheless, the full-panel disagreement is not solely one gene carrying the average. The largest full age-only absolute coefficient shares are 18.2% for fibroblast COL3A1 and 24.6% for bulk DCN. These are algebraic contribution shares, not causal fractions or variance explained. Under a shared design, the slope of the mean eight-gene score equals the mean of the eight gene slopes. The script verifies that identity within 10⁻¹². [All 192 gene effects and contributions](derived/gene_contributions.tsv) and [cross-cohort gene directions](derived/cross_cohort_gene_directions.tsv) are retained, including both partial-mixture modes and every population. ## Age-support restrictions change the answer Each cell in this table shows **age-only / adjusted** panel slopes per decade. | Population | Fibroblast only | Observed partial capture | Common-RNA partial capture | Bulk gingiva | | --- | ---: | ---: | ---: | ---: | | Original healthy cohorts | +0.172 / +0.186 | +0.213 / +0.187 | +0.183 / +0.192 | −0.145 / −0.236 | | Common observed span, 28–61 | +0.121 / +0.155 | +0.120 / +0.089 | +0.140 / +0.172 | +0.562 / +0.481 | | Shared nonempty decade bins | +0.188 / +0.189 | +0.163 / +0.136 | +0.182 / +0.180 | −0.144 / −0.237 | In the common-span bulk subset, **all eight gene slopes become positive** in both models, and the panel remains positive in every donor omission. Its adjusted maximum leverage is **0.862**, for the 28-year participant AC018. The positive narrow-range association and the negative full-range association are different sampled contrasts; they do not demonstrate a nonlinear biological trajectory or identify the reason for reversal. The shared-decade full fits retain negative bulk and positive cell-derived directions. Yet its cell-source adjusted maximum leverage is **0.724** for the single 61-year donor; bulk adjusted leverage is **0.800** for AC018. The unadjusted bulk panel changes sign when AC016, age 68, is omitted (−0.144 full versus +0.033 after omission). The adjusted bulk omission range stays negative (−0.275 to −0.131). One other fragile result stays visible: in the common-span observed partial mixture, the adjusted +0.089 slope becomes −0.082 after omitting `health:241` or −0.016 after omitting `health:148`. All the [24 panel fits](derived/fixed_panel_age_effects.tsv) retain rank, residual degrees of freedom, exact maximum-leverage identity and [252 individual donor omissions](derived/donor_omissions.tsv). Every one of the **192 single-gene omission diagnostics** retains its own full-panel sign. No seven-gene subset is selected as a replacement panel. The complete [gene-omission table](derived/all_gene_omissions.tsv) shows every omission, rather than only favorable subsets. Gene omissions are diagnostics, not eight independent biological replications. ## Research decision and limits The first question now has a bounded answer: the original full-cohort opposition is broad across this fixed ECM panel, rather than an isolated feature. The second question exposes a larger problem: changing age support can reverse the entire bulk panel, while a different coarse restriction retains the conflict. An age-signature interpretation therefore depends strongly on whom and what range was sampled. This cannot distinguish biological age dependence from cell mixture, clinical ascertainment, tissue site, sequencing/processing or other donor differences. The restrictions were frozen from metadata, but they still alter the population and do not recover absent older men or matched donor/assay pairs. The earlier common-RNA weights remain frozen from all 13 healthy cell-source donors, including donors omitted by these restrictions; this work is not held-out validation. The next discriminant remains matched-donor bulk and cell-resolved tissue with overlapping age/sex coverage, site/health qualification and independent functional measurements. Until that exists, neither increasing nor decreasing an ECM transcript mean should be treated as a rejuvenation endpoint. Dental matrix architecture, barrier, bilateral attachment, mechanical recovery and durability after finite support remain separate required outcomes. ## Reproduce and inspect ```powershell python studies/checklist-translation-2026-10-19/aging/freeze_populations.py python studies/checklist-translation-2026-10-19/aging/analyze_gene_support.py python -m unittest discover -s tests -p test_checklist_translation_aging_support.py -v ``` The analysis/tests need NumPy; demographic freezing uses the standard library. Twelve new checks pass, including fixed-gene OLS additivity, all eight omissions, exact donor sets, rank failure, coarse versus common-span support, donor influence counts and receipt hashes. The analyzer independently recomputes the healthy-only metadata populations to reject changed membership. No source acquisition, new fetch provenance, model/gene optimization or new method package was added. Docker workbench remains unavailable. [The summary](derived/analysis_summary.json) pins plan/population/script/output hashes and software; [the healthy gene table](derived/healthy_gene_values.tsv) retains the actual input values used. All previous source qualifications, unknowns and rights remain applicable. Original source data and observations belong to their authors; these are computational reanalyses and explicitly bounded diagnostics, not new wet-lab findings.