Background: The methylation status of numerous CpG sites in the human genome varies with age. The Horvath epigenetic clock used a wide variety of published DNA methylation data to produce an age prediction that has been widely used to, predict age in unknown samples, and draw conclusions about speed of ageing in various tissues, environments, and diseases. Despite its utility, there are a number of assumptions in the model that require examination. We explore the characteristics of the model in whole blood and multiple brain regions from older people, who are not well represented in the original training data, and in blood from a cross-sectional population study. Results: We find that the model systematically underestimates age in tissues from older people. A decrease in slope of the predicted ages were observed at approximately 60 years, indicating that some loci in the model may change differently with age, and that age acceleration measures will themselves be age-dependent. This is seen most strongly in the cerebellum but is also present in other examined tissues, and is consistently observed in multiple datasets. An apparent association of Alzheimer s disease with age acceleration disappears when age is used as a covariate. Association tests in the literature use a variety of methods for calculating age acceleration and often do not use age as a covariate. This is a potential cause of misleading findings. Conclusions: Associations of phenotypes with age acceleration should be evaluated cautiously, and chronological age should be included as a covariate in all analyses.

Download data

• Downloaded 789 times
• Download rankings, all-time:
  • Site-wide: 25,479
  • In genomics: 2,452
• Year to date:
  • Site-wide: 63,814
• Since beginning of last month:
  • Site-wide: 94,843

Altmetric data

Downloads over time

Distribution of downloads per paper, site-wide