A thousand people show up for an interview, but only two hundred get the job. The same goes for estimating age from DNA: people used to think you had to check a thousand sites where a chemical mark changes gene activity. Now it turns out a fifth of that is enough, and the answer barely changes.
The researchers took blood from 656 people and measured methylation at 1000 CpG loci, the ones most strongly linked to age. A CpG locus is a spot in DNA where a cytosine sits right before a guanine. Then they ran a program that picks which sites to keep on its own while simultaneously tuning the prediction algorithm. The program kept an average of 212 sites, and the age prediction error grew by only 20 days. For comparison, even repeated measurements on the same data make the error jump by 0.3 years, five times more.
So you can make cheaper biological age tests by checking fewer spots. But if some site seems biologically important, that has to be checked separately, because statistics does not guarantee that this particular site is what affects aging.
The error difference between the full model and MOSurvivor was 0.054 years, which is about 19.8 days. For comparison, the standard deviation of the error across different data splits was around 0.3 years, five times larger. So saving 788 CpGs costs less than the natural spread that comes from choosing a training set.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.