DeepMAge: A Methylation Aging Clock Developed with Deep Learning

Abstract

DNA methylation aging clocks have become an invaluable tool in biogerontology research since their inception in 2013. Today, a variety of machine learning approaches have been tested for the purpose of predicting human age based on molecular-level features. Among these, deep learning, or neural networks, is an especially promising approach that has been used to construct accurate clocks using blood biochemistry, transcriptomics, and microbiomics data-feats unachieved by other algorithms. In this article, we explore how deep learning performs in a DNA methylation setting and compare it to the current industry standard-the 353 CpG clock published in 2013. The aging clock we are presenting (DeepMAge) is a neural network regressor trained on 4,930 blood DNA methylation profiles from 17 studies. Its absolute median error was 2.77 years in an independent verification set of 1,293 samples from 15 studies. DeepMAge shows biological relevance by assigning a higher predicted age to people with various health-related conditions, such as ovarian cancer, irritable bowel diseases, and multiple sclerosis.

Keywords: DNA methylation; aging; artificial intelligence; epigenetics.

Figure 1.

Scatter plot of DeepMAge predictions in 4 data cohorts”. DeepMAge accurately predicted the chronological age of healthy people from the training set (A), healthy people from the verification set (B), and remained accurate in the aggregations of case cohorts from the studies included in the training set (C) and the verification set (D). Scatter plot in panel A shows the per-fold predictions obtained during CV, and the other panels show the predictions by the final model. MedAE = Median absolute error measured in years, N = Number of samples in a corresponding cohort (see Supplementary Figures 1-3 for a more detailed visualization).

Figure 2.

The DeepMAge prediction age distribution in the verification set closely resembled the real age distribution. Distributions were obtained using Gaussian kernel with 0.3σ bandwidth, where σ is the standard deviation of the age values.

Figure 3.

DeepMAge, but not the 353 CpG clock, predicted donors with IBD (GEO study accession GSE87640) to be on average 1.23 years older than the healthy donors from the same study (p-value = 1.24E­3). Outliers outside the (-20; +20) prediction error window were removed from the image; The box is formed by the interquartile range with the median marked inside it. Whiskers protrude no farther than 1.5 times the interquartile range. GEO = Gene Expression Omnibus; IBD = Inflammatory bowel disease; N= Number of samples in a corresponding cohort.

Figure 4.

The DeepMAge clock shares 122 CpGs with the 353 CpG clock and seven CpGs with the 71 CpG clock. The latter two were published in 2013.