DeepMAge — in defense of first-generation aging clocks
Recently Deep Longevity, a longevity medicine startup based in Hong Kong, has published an article in “Aging and Disease”, in which they describe a new aging clock — DeepMAge. It is a DNA methylation (DNAm) aging clock, but unlike its predecessors which were trained using linear algorithms, DeepMAge is a neural network. The authors suggest that their model choice is what let them achieve an unprecedented age prediction accuracy of 2.77 years.
In 2013 two seminal papers on the epigenetics of aging were published: Horvath’s aging clock and Hannum’s aging clock [1, 2]. These articles present linear regression models that sum methylation levels at specific CpG loci and produce an estimation of chronological age.
The “aging clock” concept has quickly permeated biogerontology. Many more human and murine aging clocks have been built since then, most of which use DNAm information. However, some aging clocks also can interpret blood biochemistry parameters, gene expression levels, and even gut flora diversity in the context of aging [3–5].
Such models allow researchers to test hypotheses on geroprotective interventions, discover anti-aging compounds, and find connections between age and disease. Their utility cannot be underestimated in the modern world whose population is getting progressively older.
By 2050 the global median age is expected to reach 36 years, which is 6 years older than today [6]. For comparison, the global median age was 24 years in 1950. Population aging puts pressure on the healthcare and welfare systems, ultimately hurting the world economy. To prevent demographic and economic collapse, it essential to develop, test, and distribute geroprotective therapies that would let the bulk of the human population remain productive members of society significantly longer.
Aging clocks are the tools that would let researchers measure aging as a biological process with a single snapshot. They decrease the horizon of human longitudinal studies from decades to months, which is an absolute must when clinically tested geroprotectors need to be created within a single human lifetime.
The most recent development in the field of aging quantification, or “biohorology” (from Greek bio- “life” and hora- “time”) is the time-to-death clocks. Frequently called “second generation” clocks, they are trained on mortality linked data. DNAm clocks such as PhenoAge and GrimAge have recently been shown to be associated with a variety of aging outcomes, unlike the first generation DNAm clocks [7, 8].
It should be noted, however, that second-generation clocks by design require chronological age correction, and thus, their mortality estimate is not derived from purely biological information. This circumstance means that the conventional, first-generation aging clocks should not be discarded as obsolete.
Besides, there are strong indications that there is still much room for their improvement. DeepMAge is the most recent DNAm aging clock that technically belongs to the first generation of aging clocks. It is, however, quite different from its alternatives [9].
Unlike any other DNAm clocks, DeepMAge was trained as a deep neural network. This type of regressor can use non-linear variable dependencies, which would remain unseen to the linear models, such as elastic net used in most other DNAm clocks. This model choice let DeepMAge achieve a record low error rate: a median absolute error of 2.77 years in a collection of 15 independent studies and 1,293 healthy people. Even when the original elastic net approach was reproduced from scratch, the linear model achieved a significantly higher median absolute error of 3.23 years.
DeepMAge was also verified in a set of case-control studies and displayed significantly higher prediction errors in cases of obesity, ovarian cancer, frontotemporal dementia, multiple sclerosis, and inflammatory bowel diseases.
DeepMAge has not yet been tested as a mortality risk factor, although it is intriguing to see whether it could be associated with aging outcomes.
Deep neural networks are also much more flexible than linear models. The abundance of different architectures allows biohorologists to experiment with generative and variational networks to emulate geroprotective interventions in silico. At its current state, this technology can be used to streamline target and hit optimization in the pharmacological development pipeline. But ultimately, it has the potential to completely remove the need for model and human subjects [10].
Conclusion
Second generation aging clocks are invaluable tools and a logical extension of the original concept. The very term “second generation” implies that the idea of a chronological age regressor is outdated and obsolete. Meanwhile, DeepMAge demonstrates that, on the contrary, there is still much to explore within this methodology. The synthesis of biohorology and deep learning opens the path to a multitude of applications.
[1] Horvath S (2013). DNA methylation age of human tissues and cell types. Genome Biol, 14:R115–R115.
[2] Hannum G, Guinney J, Zhao L, Zhang L, Hughes G, Sadda S, et al. (2013). Genome-wide Methylation Profiles Reveal Quantitative Views of Human Aging Rates. Mol Cell, 49:359–367.
[3] Mamoshina P, Kochetov K, Putin E, Cortese F, Aliper A, Lee W-S, et al. (2018). Population Specific Biomarkers of Human Aging: A Big Data Study Using South Korean, Canadian, and Eastern European Patient Populations. J Gerontol Ser A, 73:1482–1490.
[4] Mamoshina P, Volosnikova M, Ozerov IV, Putin E, Skibina E, Cortese F, et al. (2018). Machine Learning on Human Muscle Transcriptomic Data for Biomarker Discovery and Tissue-Specific Drug Target Identification. Front Genet, 9:242.
[5] Galkin F, Mamoshina P, Aliper A, Putin E, Moskalev V, Gladyshev VN, et al. (2020). Human Gut Microbiome Aging Clock Based on Taxonomic Profiling and Deep Learning. iScience, 23:101199.
[6] Ritchie H (2019). Age Structure. Our World Data .
[7] Lu AT, Quach A, Wilson JG, Reiner AP, Aviv A, Raj K, et al. (2019). DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging, 11:303–327.
[8] Levine ME, Lu AT, Quach A, Chen BH, Assimes TL, Bandinelli S, et al. (2018). An epigenetic biomarker of aging for lifespan and healthspan. Aging, 10:573–591.
[9] Fedor Galkin PM, Fedor Galkin PM DeepMAge: A Methylation Aging Clock Developed with Deep Learning. Aging Dis, 0.
[10] Galkin F, Mamoshina P, Aliper A, de Magalhães JP, Gladyshev VN, Zhavoronkov A (2020). Biohorology and biomarkers of aging: Current state-of-the-art, challenges and opportunities. Ageing Res Rev. doi: 10.1016/j.arr.2020.101050.
