Epigenetic aging clocks are built on hundreds of thousands of CpG loci, yet only a small fraction of them carry age-related information. Whole-genome arrays are expensive and redundant for routine screening. This study sits alongside research on signal filtering and feature selection, where compact representations are also sought. The paper itself is somewhat of an outlier: its closest neighbors deal with neural-network detectors of antibiotic resistance, maps of cellular senescence, and gene-based predictors of breast cancer. The broader thematic group is tied to Fourier transforms and signal processing, but there is no direct overlap in methods. The work answers a practical question: can you keep the accuracy of a clock while retaining only a small share of its markers?
The authors used the public GSE40279 dataset: DNA methylation in whole blood from 656 people of various ages, measured with the Illumina 450K chip. For the initial selection, they took 1000 CpGs with the highest absolute correlation with age. The data were then randomly split 30 times into training (80%) and test (20%) sets using different random seeds. Each candidate in MOSurvivor encoded weights for all 1000 CpGs, a selection threshold, and eight XGBoost hyperparameters (tree depth, learning rate, number of trees, row and column fractions, minimum leaf weight, gamma, L2 regularization). A CpG was activated if its weight exceeded the threshold; a repair operator kept the number of selected loci within the range of 50-300. Optimization followed a survival-of-the-fittest principle: the population was divided into two groups, pairwise competitions determined the winners, and candidate updates took into account the best in each pair, the best in each group, the global leader, and leaders from the Pareto archive. The final solution was chosen by the sum of the cross-validation mean absolute error and a penalty for subset size (0.10 × the fraction of selected CpGs). For comparison, the authors used the full model on all 1000 CpGs, random search, a genetic algorithm, and particle swarm optimization.
The full XGBoost model on all 1000 CpGs showed the lowest mean absolute error (MAE) at 4.095 years. MOSurvivor had an MAE of 4.149 years, PSO 4.124, the genetic algorithm 4.170, and random search 4.197. At the same time, MOSurvivor selected an average of 211.6 CpGs versus 293.4 for PSO, 269.6 for GA, and 278.7 for random search. Relative to 1000 CpGs, that is a dimensionality reduction of 78.8% at the cost of an error increase of 0.054 years (roughly 19.8 days). Paired Wilcoxon tests with Holm correction revealed no statistically significant differences with any of the competitors (all p > 0.05). MOSurvivor was more accurate than the full model on only 10 of the 30 splits. The stability of the selection turned out to be low: the average Jaccard index between pairs of the 30 subsets was 0.124 ± 0.030. Nine CpGs appeared in at least 15 runs, but none showed up in all 30. The most frequent locus, cg16867657, appeared in 29 runs, followed by cg04400972 (24) and cg14361627 (20).
The work shows that jointly optimizing feature selection and hyperparameters can compress an epigenetic clock fivefold without a statistically significant loss of accuracy. Unlike neighboring studies that emphasize absolute accuracy or biological interpretation, this one tackles a computational problem: finding a minimal set of markers. That is a different scale entirely, not the discovery of new biomarkers but engineering resource savings. The low reproducibility of selection across runs points to high redundancy among CpGs: different sets yield nearly identical predictions. In practice, this means cheaper targeted panels can be designed, but the biological significance of individual loci should be verified separately.
A promising next step is to embed CpG selection into an outer cross-validation loop to eliminate information leakage from the test data. Biological annotation of the selected loci could also be added, along with checks for their enrichment in genes and regulatory regions. Comparing against the published Horvath and Hannum clocks on the same samples would show how much a compact model falls short of, or matches, the classics. The approach could be extended to other tissues and cohorts, and adapted to other types of omics data where features are many and samples are few.
The results will be useful to developers of epigenetic aging tests: smaller panels are cheaper to produce and easier to interpret. The joint optimization method could be transferred to other high-dimensional tasks, such as selecting transcriptomic or proteomic markers.
Repeat the analysis with proper CpG selection inside each training split and carry out external validation on an independent cohort. Equalize the computational budgets for all optimizers and report runtimes so that efficiency can be judged.
The key unsolved problem in the epigenetics of aging is the biological interpretability of clocks. The low reproducibility of CpG selection in this work highlights that statistically optimal sets may not coincide with causally significant loci. The link between methylation at specific CpGs and the mechanisms of aging remains an open question, and this paper does not bring us closer to answering it; it merely shows that prediction can be achieved through many equivalent paths.
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.