Epigenetic age-predictor for mice based on three CpG sites

Age-associated DNA methylation (DNAm) was first described for humans after Illumina Bead Chip microarray data became available to enable cross comparison of thousands of CpG loci (Bocklandt et al., 2011; Koch and Wagner, 2011). Many of these age-associated CpGs were then integrated into epigenetic age-predictors (Hannum et al., 2013; Horvath, 2013; Weidner et al., 2014). However, site-specific DNAm analysis at individual CpGs can also provide robust biomarkers for aging. For example, we have described that DNAm analysis at only three CpGs enables age-predictions for human blood samples with a mean absolute deviation (MAD) from chronological age of less than five years (Weidner et al., 2014). Such simplistic age-predictors for human specimen are widely used because they enable fast and cost-effective analysis in large cohorts.

Recently, epigenetic clocks were also published for mice by using either reduced representation bisulfite sequencing (RRBS) or whole genome bisulfite sequencing (WGBS) (Petkovich et al., 2017; Stubbs et al., 2017; Wang et al., 2017). For example, Petkovich et al. described a 90 CpG model for blood (Petkovich et al., 2017), and Stubbs and coworkers a 329 CpG model for various different tissues (Stubbs et al., 2017). Nutrition and genetic background seem to affect the epigenetic age of mice – and thereby possibly aging of the organism (Cole et al., 2017; Hahn et al., 2017; Maegawa et al., 2017). In analogy, epigenetic aging of humans is associated with life expectancy, indicating that it rather reflects biological age than chronological age (Lin et al., 2016; Marioni et al., 2015). However, DNAm profiling by deep sequencing technology is technically still challenging, relatively expensive, and not every sequencing-run covers all relevant CpG sites with enough reading depth.

Therefore, we established pyrosequencing assays for nine genomic regions of previously published predictors (Petkovich et al., 2017; Stubbs et al., 2017). These regions were preselected to have multiple age-associated CpGs in close vicinity. DNAm was then analyzed in 24 blood samples of female C57BL/6 mice that covered a broad range of 12 different age groups (11 to 117 weeks old). The nine amplicons covered a total of 71 CpG sites (Supplementary file 1) and we used machine learning to identify the best fitted model for epigenetic age-predictions using cross-fold validation on the training set. The best results were observed for 15 CpGs from five different amplicons that provided an extremely high correlation with chronological age in the training set (R² = 0.99; mean absolute deviation [MAD]=2.76 weeks; Supplementary file 2), albeit the training set might be too small for this approach. To make the method more easily applicable and more cost-effective, we wanted to focus on less CpGs. When we varied the regularization parameters for models with less CpGs, the precision declined significantly. For example the best model with three CpGs comprised the three CpGs of Hsf4 (CpGs# 3,4,5) that also revealed the overall highest Pearson correlations with chronological age (R² = 0.95; MAD = 5.24 weeks). However, combination of different hypo- and hypermethylated amplicons might be advantageous to facilitate better assessment of plausibility of the results. Therefore, we alternatively selected those three CpGs that revealed the highest Pearson correlation with chronological age in different amplicons. These three CpGs were associated with the genes Proline rich membrane anchor 1 (Prima1: chr12:103214639; R² = 0.71), Heat shock transcription factor 4 (Hsf4: chr8:105271000; R² = 0.95) and Potassium voltage-gated channel modifier subfamily S member 1 (Kcns1: chr2:164168110; R² = 0.83; Figure 1A–C; Figure 1—figure supplement 1). Notably, all three CpGs were derived from the epigenetic age-predictor for blood samples (Petkovich et al., 2017). A multivariable model for age-predictions was established for DNAm at the CpGs in Prima 1 (α), Hsf4 (β), and Kcns1 (γ):

Predicted age^C57BL/6 (in weeks) = −58.076 + 0.25788 α + 3.06845 β + 1.00879 γ

Age-predictions correlated very well with the chronological age of C57BL/6 mice in the training set (R² = 0.96; MAD = 4.86 weeks; Figure 1D).

Our three CpG age-predictor was subsequently validated in a blinded manner for 21 C57BL/6J mice (7 to 104 weeks old) from the University of Ulm (validation set 1) and 19 C57BL/6J mice (14 to 109 weeks old) from the University of Groningen (validation set 2). The results of both validation sets revealed high correlations with chronological age (R² = 0.95 and 0.91, respectively; Figure 1E–H) with relatively small MADs (6.9 and 7.1 weeks) and median absolute errors (MAE; 5.0 and 5.9 weeks). Thus, our age-predictions seem to have similar precision as previously described for multi-CpG predictors based on RRBS or WGBS data (Petkovich et al., 2017; Stubbs et al., 2017; Wang et al., 2017).

Gender did not have significant impact on our epigenetic age-predictions for mice (Figure 2), as described before (Maegawa et al., 2017; Petkovich et al., 2017; Stubbs et al., 2017). In contrast, the human epigenetic clock is clearly accelerated in male donors (Hannum et al., 2013; Horvath, 2013; Weidner et al., 2014). This coincides with shorter life expectancy in men than woman, whereas in mice there are no consistent sex differences in longevity (Goodrick, 1975).

To address the question if our three CpG signature was also applicable for other tissues than blood we analyzed the DNAm in skin, kidney, intestine, lung, liver, heart, brain, testis, and pancreas of 3 young (9.6 weeks old) and three old mice (56.9 weeks old). In all tissues tested the samples of old mice were predicted to be older using our three CpG signature. However, the different DNAm levels clearly demonstrate that the model needs to be retrained to be applied for these tissues (Figure 3).

Subsequently, we analyzed epigenetic aging of DBA/2 mice that have a shorter life expectancy than C57BL/6 mice (Goodrick, 1975) (33 mice from Ulm and Groningen; 6 to 109 weeks old). The three CpGs in Prima1, Hsf4 and Kcns1 revealed high correlation with chronological age (R² = 0.91, 0.88 and 0.83, respectively), albeit the offset in DNAm between DBA/2 and C57BL/6 mice indicated that the signature needs to be retrained for different mouse strains (Figure 4a–c). Notably, the slopes were higher in DBA/2 mice, particularly for the CpG in Prima1. Furthermore, DNAm of Hsf4 increased at a higher rate in young DBA/2 mice, indicating that it is more accurately modelled as a function of logarithmic age. This has also been described in human for many age-associated CpGs in pediatric cohorts (Alisch et al., 2012). In fact, epigenetic age-predictions in DBA/2 mice seemed to follow a logarithmic model of age (R² = 0.89; Figure 4d) rather than a linear association (R² = 0.86). These results provided evidence for accelerated epigenetic aging of DBA/2 mice.

Either way, epigenetic age-predictions were overall significantly overestimated in the shorter-lived DBA/2 mice, suggesting that age-predictors need to be adjusted for different inbreed mice strains. To this end, we have retrained a multivariate model for DBA/2 mice:

Predicted age^DBA/2 (in weeks) = 87.54294–1.22221 α + 0.991558 β + 0.355444 γ

This adjusted model facilitated relatively precise age-predictions for DBA/2 mice (R² = 0.95; MAD = 7.1 weeks; MAE = 5.3 weeks; Figure 4e).

Generation of confined epigenetic signatures is always a tradeoff between integrating more CpGs for higher precision and higher costs for analysis (Wagner, 2017). It was somewhat unexpected that with only three CpGs our signature facilitated similar precision of epigenetic age-predictions as the previously published signatures based on more than 90 CpGs. This can be attributed to the higher precision of DNAm measurements at individual CpGs by bisulfite pyrosequencing, which is one of the most precise methods for determining DNAm at single CpG resolution (BLUEPRINT consortium, 2016). Particularly in RRBS data not all CpG sites are covered in all samples and a limited number of reads notoriously entails lower precision of DNAm levels at these genomic locations. Thus, genome wide deep sequencing approaches facilitate generation of robust large epigenetic age-predictors, while site specific analysis may compensate by higher precision of DNAm measurement at individual CpGs.

The ultimate goal of epigenetic age-predictors for mice is not to develop near perfect age predictors, but to provide a surrogate for biological aging that facilitates assessment of interventions on aging. In fact, using deep sequencing approaches (RRBS or WGBS) several groups already indicated that relevant parameters that affect aging of the organism - such as diet, genetic background, and drugs - do also impact on epigenetic aging (Cole et al., 2017; Hahn et al., 2017; Maegawa et al., 2017). It is yet unclear if epigenetic aging signatures can be specifically trained to either correlate with chronological age or biological age. For humans, recent studies indicate that this might be possible (Levine et al., 2018) and we have previously demonstrated that even individual age-associated CpGs can be indicative for life expectancy (Zhang et al., 2017). Further studies will be necessary to gain better understanding how epigenetic age predictions are related to the real state of biological aging, and how it is related to alternative approaches to quantify biological aging, such as telomere length (Belsky et al., 2018).

Our three CpG model has been trained for blood samples – a specimen that is commonly used in biochemical analysis and the small required volume can be taken without sacrificing the mice. However, epigenetic aging may occur at different rates in different tissues. It is difficult to address this question in humans because it is difficult to collect samples of various tissues in large aging cohorts, whereas this is feasible in mice. We demonstrate that age-associated DNAm changes occur in multiple tissues in our three CpGs albeit they were initially identified in blood (Petkovich et al., 2017). Furthermore, DNAm levels may vary between different hematopoietic subsets (Frobel et al., 2018; Houseman et al., 2014). In the future, sorted subsets should be analyzed to determine how the three CpG signature is affected by blood counts.

The results of our three CpG signature suggest that epigenetic aging is accelerated in DBA/2 mice. Notably, in elderly DBA/2 mice the epigenetic age predictions revealed higher ‘errors’ from chronological age, which might be attributed to the fact that the variation of lifespan is higher in DBA/2 than C57BL/6 mice (de Haan et al., 1998; Goodrick, 1975). It will be important to validate the association of the epigenetic age-predictions with biological age by additional correlative studies, including life expectancy in mice.

Taken together, we describe an easily applicable but quite precise approach to determine epigenetic age of mice. We believe that our assay will be instrumental to gain additional insight into mechanisms that regulate age-associated DNAm and for longevity intervention studies in mice.

This work was supported by the Else Kröner-Fresenius-Stiftung (2014_A193; to WW), by the German Research Foundation (DFG; WA 1706/8-1 to WW; and GRK 1789 CEMMA, GRK 2254 HEIST and SFBs 1074, 1149 and 1275 to HG), by the German Ministry of Education and Research (BMBF; 01KU1402B to WW; and SyStarR to HG), and by the NIH (R01HL134617 and R01DK104814 to HG). The Groningen samples were obtained from the Mouse Clinic for Cancer and Ageing (http://www.mccanet.nl↗), which is supported by a grant from the Netherlands Organization for Scientific Research (NWO). The funding bodies were not involved in study design, data analysis, or writing of the manuscript.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Wolfgang Wagner, Email: wwagner@ukaachen.de.

Vadim N Gladyshev, Brigham and Women's Hospital, Harvard Medical School, United States.

Jessica K Tyler, Weill Cornell Medicine, United States.

This paper was supported by the following grants: