The Evidence Layer for Scientific AI

 Quick Analysis Plan Copied

 Long Analysis Plan

Core evidence: the eFRS

The anchor study profiles blood DNA methylation (EPIC/450K arrays) in German cohorts: discovery in ESTHER subset I (n=998) found 2,220 CpGs at FDR<0.05, of which 65 replicated in subset II (n=730), and LASSO selected 20 CpGs into the eFRS. eFRS–FI correlations were only ~0.2–0.3, and eFRS explained ~6–9% of FI variance eFRS Nature Communications study.

The decisive weakness: adding eFRS to age+sex improved prediction significantly in ESTHER (p=0.01) but not in KORA-Age (p=0.23), an internal replication failure for incremental value eFRS incremental AUC results. Most signal likely reflects accelerated epigenetic aging shared with GrimAge and residual smoking confounding.

Context: what methylation captures in frailty

A smoking-methylation ensemble in HIV veterans predicted frailty (AUC 0.78 replication) far better than self-reported smoking status (AUC 0.55), showing methylation indices encode exposures and inflammation relevant to frailty Smoking methylation HIV frailty ML study, and meta-analysis shows CRP/IL-6 associate cross-sectionally with frailty but fail to predict onset longitudinally Inflammation-frailty meta-analysis — a warning that cross-sectional methylation–frailty associations may not be predictive.

Battle-test verdict and falsification path

Confidence: moderate. Blindspots: German-only cohorts (no ancestry generalization), array-restricted CpG coverage, self-reported FI deficits, health-check selection bias. The eFRS would be falsified if it shows no FI association after age/sex adjustment, no replication, or no AUC gain over age/sex across multiple external populations — the KORA-Age result already partially satisfies the last criterion. Whether eFRS captures causal biology or is a surrogate for smoking/inflammation remains unresolved.