A thousand candidates are sitting in the interview room, but only two hundred will be hired. It used to seem that to accurately estimate age from DNA, you had to interview every one of the thousand sites where a chemical mark alters gene activity. Now it's clear: you can hear out just a fifth of them, and the final decision barely changes.
The researchers took blood from 656 people and measured methylation at 1000 CpG loci most strongly associated with age. They then built the MOSurvivor optimizer, which simultaneously selects informative sites and tunes eight parameters of the XGBoost algorithm (a tree-based machine learning method). Each candidate in this optimizer encoded weights for all 1000 sites, the selection threshold, and model settings. A site was activated if its weight exceeded the threshold, and a special operator kept the number of selected loci between 50 and 300.
As a result, MOSurvivor kept an average of 211.6 CpG loci, reducing dimensionality by 78.8% compared with the full model. The mean absolute error in age prediction rose by only 0.054 years, roughly 20 days. For comparison, the natural spread of error across different data splits is about 0.3 years, five times larger. Statistical tests found no significant differences from the full model or other optimizers.
But here's the curious part: if you run the selection again, the list of 212 sites barely overlaps with the previous one. The average Jaccard index (a measure of similarity between two sets) between pairs of subsets was just 0.124. This means CpG loci are highly redundant: different sets yield almost the same prediction. In practice, this opens the door to cheap targeted panels where saving on the number of probes matters, but the biological significance of individual sites will have to be checked separately.
So an interview that picks two hundred out of a thousand candidates without questioning the rest gives almost the same result. The only question is who exactly makes it onto that short list, and next time it could be completely different.
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.