Research note | 16 November 2026 | The Null Institute | Summary-data reanalysis
Our result: Across five interventions reported in a 2026 Frontiers in Aging study, the number of female-mouse tissues with a statistically significant model-score difference decreased rather than increased as the intervention’s published female lifespan gain grew. Our treatment-level recalculation gives Spearman’s ρ = −0.789; an exact two-sided permutation test gives p = 0.15. This is an observed pattern in five interventions, not statistically compelling evidence of an inverse relationship.
The study tested whether metabolomic patterns can flag treatments associated with longer life. Its male-mouse results distinguished treated animals from controls in 33 of 35 tissue–intervention comparisons. Our calculation concerns the paper’s separate female-mouse analysis, not the male finding. Read the original study by Badenoch and colleagues.
The five interventions
| Intervention | Published female median lifespan gain | Significant tissue comparisons |
|---|---|---|
| Caloric restriction | +40% | 1/7 |
| Rapamycin | +12.6% | 2/7 |
| Acarbose | +5% | 1/7 |
| Canagliflozin | 0%* | 6/7 |
| 17α-estradiol | 0%* | 3/7 |
*For canagliflozin and 17α-estradiol, the study assigned a 0% female-lifespan benefit because earlier studies had not found extension. None of these lifespan changes was measured by waiting for the female mice in this metabolomics cohort to die. The study reports the seven-tissue significance counts used here.
What the recalculation adds
We treated each intervention as one observation (n = 5), rather than pretending that 35 related tissue tests were independent biological interventions. We ranked the five published lifespan-gain labels against the counts of significantly separated tissues and obtained ρ = −0.78947. Enumerating all 60 unique permutations of the five count labels gives 9 correlations at least as extreme in absolute value: exact two-sided p = 9/60 = 0.15.
For a more intuitive descriptive comparison, the two interventions without a reported female lifespan benefit were distinguishable in 9/14 tissue comparisons (64.3%), while the three with reported lifespan benefits were distinguishable in 4/21 (19.0%). That is a 3.38-fold difference in observed frequencies, not a 3.38-fold difference in lifespan-prediction accuracy. The tissue tests overlap biologically and statistically, so this ratio is not an independent-trials estimate.
Detecting treatment exposure versus predicting longevity
A significantly different model score means the model detected a difference between treated animals and controls. It does not, by itself, show that the score correctly predicted the direction or amount of a lifespan change. This matters when drugs alter metabolism without extending lifespan in a particular sex. The original authors already acknowledge that their female model estimates did not track group-specific lifespan. Our contribution is the explicit five-intervention correlation and exact permutation check.
What we could not test: A direct error metric such as mean absolute error requires numeric intervention-level model predictions, preferably individual-animal values, not just counts of tissues passing significance thresholds. The article reports counts and references Supplementary Figure S1, but a machine-readable table of female predictions was not verified in this reanalysis. We did not infer numerical error rates by visually estimating a plot.
Reproduce the calculation
The entire intervention-level input is the five rows above. With SciPy installed, this code prints the correlation and exact permutation p-value:
from itertools import permutations
from scipy.stats import spearmanr
lifespan = [40, 12.6, 5, 0, 0]
detected_tissues = [1, 2, 1, 6, 3]
observed = spearmanr(lifespan, detected_tissues).statistic
null = [spearmanr(lifespan, p).statistic
for p in set(permutations(detected_tissues))]
extreme = sum(abs(r) >= abs(observed) - 1e-12 for r in null)
print(round(observed, 6), f"{extreme}/{len(null)}", extreme / len(null))
# -0.789474 9/60 0.15
Limits: five interventions; two intervention-level lifespan labels coded as zero; a significance-count proxy rather than raw predictions; correlated tissues; and group-level survival estimates from different cohorts. This is a reproducible summary-data check, not a rerun of the original machine-learning model. It does not establish that the model is useless or that metabolic changes cannot eventually predict longevity.
Source: Badenoch B, Fiehn O, Rappaport N, Greenfield S, Chandrasekaran S, Miller RA. “Multi-tissue metabolomic signatures of five longevity interventions converge on ergothioneine and lipid remodeling in male UM-HET3 mice.” Frontiers in Aging, 6 November 2026. DOI: 10.3389/fragi.2026.1972744 (female analysis described in Methods and Discussion; Supplementary Figure S1).
AI-assisted drafting and computation; source figures and method were checked against the cited study and by rerunning the included code. No external peer review. The Null Institute, 2026.