Albert Ying's Clockbase as the *best* epigenetic age tracker

All the other epigenetic clocks have their issues (the 1st and 2nd generation clicks do not respond to calorie restriction, and even DunedinPACE increases after partial reprogramming). The meaningfulness of any of the CpG sites is unknown (and are not significant to the calculation of DamageAgea and AdaptAge). But the DamageAge and AdaptAge Ying CpG sites are different and the real ones

https://www.researchgate.net/publication/368982451_ClockBase_a_comprehensive_platform_for_biological_age_profiling_in_human_and_mouse
https://www.researchgate.net/publication/364266802_Causal_Epigenetic_Age_Uncouples_Damage_and_Adaptation

4 Likes

I got my data plugged in and there are VERY VERY sharp discrepancies that I don’t know what to make of yet…

DunedinPACE is 0.59. Of all the epigenetic clocks, PhenoAge is by far the lowest (the first-generation epigenetic clocks have little functional significance and I can cry less over them b/c I do way better on 2nd/3rd generation). The first-generation epigenetic clocks are MORE correlated with AdaptAge than DamageAge (meaning HorvathAge and especially HannumAge are more associated with “adaptive/protective” changes w/age than damaging changes w/age).

But DunedinPoAM (precursor and less accurate than DunedinPACE) is barely over 1.

DamageAge and AdaptAge… I don’t know what to make of yet… There’s a huge discrepancy I need to figure out better, because there’s a chance this massive discrepancy could be rare enough to be socially important…

1 Like

HannumAge is MORE associated with adapt-age than damage-age, as is HorvathAge. DNAm that contribute to HorvathAge/HannumAge have little functional significance (in fact, slight acceleration of HannumAge may be protective - HannumAgeAA has the opposite effects of GrimAgeAA)

GrimAge outperforms phenoage…

Then there’s GrimAge2, which supposedly outperforms GrimAge (GrimAge2 is ONLY trained on people over 40 )

To arrive at DNA methylation based surrogates of these plasma proteins, we used two elastic net regression models to predict log-transformed (base e) versions of high-sensitivity C-reactive protein (log CRP) and hemoglobin A1C (log A1C), respectively. Both elastic net regression models used the following candidate covariates: 1030 CpGs, Age and Female. The two elastic net regression models selected 132 CpGs (for log CRP) and 86 CpGs (for log A1C), respectively (Supplementary Table 1). The predicted values resulting from these regression models will be denoted by DNAm logCRP and DNAm logA1C, respectively. The Pearson correlation coefficients between the DNAm variables and their target proteins are biased in the training dataset (Supplementary Figure 1A, 1B) due to overfitting. Our unbiased analysis in the test dataset leads to the following: Pearson correlations r=0.48 for DNAm logCRP and r=0.34 for DNAm logA1C (Supplementary Figure 1C, 1D).

To define GrimAge2 we used a Cox regression model to regress time-to-death (due to all-cause mortality) on the following candidate covariates: eleven DNAm-based surrogates of plasma proteins, DNAm PACKYRS, Age, Female (Methods, Supplementary Table 1). We remind the readers that the first version of GrimAge was based on Age, Female, DNAm PACKYRS, and seven DNAm-based proteins: adrenomedullin (ADM), beta-2-microglobulim (B2M), cystatin C (Cystatin C), GDF-15, leptin (Leptin), PAI-1, and tissue inhibitor metalloproteinases 1 (TIMP-1, Supplementary Note 1). Interestingly, the Cox regression model with a elastic net penalty picked up the exactly same seven DNAm proteins, DNAm PACKYRS, as well as the two new biomarkers (DNAm logCRP and DNAm logA1C). Thus, GrimAge2 is based on 12 covariates: 10 DNAm based biomarkers and 2 demographic characteristics: Age, Female (Figure 1). The linear combination of covariates resulting from the elastic net Cox regression model can be interpreted as an estimate of the logarithm of the hazard ratio of mortality. We calibrated this parameter into an age estimate by performing a linear transformation whose slope and intercept terms were chosen by forcing the mean and variance of DNAm GrimAge2 to match that of chronological age in the training data (Figure 1).

Talked to jesse pogalnik yesterday, who has some experience with it (and who recently published the DamageAge/AdaptAge papers). There’s thought that AdaptAge increases after recovery from stress [stress is when DamageAge may increase]. We still don’t know if DamageAge/AdaptAge are more dynamic/changeable/intervention-friendly than other clocks

1 Like