Ученые применили новый алгоритм машинного обучения для проверки влияния зоны долголетия Никаой на эпигенетическое старение

A new machine-learning framework puts one of the world’s best-known blue zones to the test, asking whether Nicoya’s exceptional longevity leaves a measurable signature in DNA methylation. Study: Aging Out of the Blue: Region-Specific Epigenetic Clock Calibration for a Blue Zone with the DNAm SuperLearner. Image Credit: Denis Kvarda / Shutterstock *Important notice: medRxiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information. In a recent study posted on the medRxiv preprint* server, researchers presented a novel framework that combines established epigenetic clocks with flexible machine-learning algorithms within SuperLearner to predict and calibrate epigenetic aging. The human methylome is an additional layer of genome regulation, in which methyl groups attach to cytosine-phosphate-guanine (CpG) sites, modulating gene expression and other cellular processes. Epigenetic clocks estimate biological age based on DNA methylation (DNAm) patterns at CpG sites. DNAm-based age estimators are among the most accurate measures of biological aging. Blue zones are regions characterized by exceptionally high numbers of centenarians, offering a unique opportunity to study how biological aging diverges from chronological age. Nevertheless, standard epigenetic clocks trained on large, heterogeneous populations capture global average age–methylation associations rather than region-specific dynamics. About the study In the present study, researchers described a SuperLearner framework for predicting epigenetic age. They used data from the Costa Rican Longevity and Healthy Aging study, which recruited 2,827 Costa Ricans aged ≥ 60 years in 2004–06. From this cohort, 1,081 individuals provided blood samples for DNAm profiling. This sample comprised 875 participants from elsewhere in Costa Rica and 206 residents of the Nicoya Peninsula, one of the world’s best-documented blue zones, with exceptionally long-lived Nicoyans aged ≥ 95 years oversampled. A SuperLearner, an ensemble machine learning (ML) algorithm, was trained on non-Nicoyans as the reference (comparison) group. The researchers constructed the SuperLearner ensemble from a library of candidate learners spanning both ML algorithms and fixed epigenetic clocks. DNAm data were split into training folds. Epigenetic clocks and ML algorithms were fit on training folds and subsequently used to predict epigenetic age on the validation (left-out) set. Next, a risk function (i.e., mean squared error, MSE) was computed iteratively on the validation set for each algorithm, yielding cross-validated risk estimates. These cross-validated predictions were then combined, with greater weight assigned to models that predicted age more accurately. The resultant weighted combination of algorithms constituted the SuperLearner ensemble, while the candidate learner with the lowest cross-validated risk was termed the discrete SuperLearner. Findings The SuperLearner ensemble comprised four epigenetic clocks, principal component GrimAge (PCGrimAge), elastic net (EN), EN centenarian 40 (ENCen40), and Hannum clocks, and one ML algorithm (ridge regression). The PCGrimAge clock accounted for 73.6% of the ensemble weight, while the Hannum, EN, and ENCen40 clocks accounted for 16.1%, 7.9%, and 1.7%, respectively. The ML algorithm had only 0.7% of the weight. PCGrimAge emerged as the discrete SuperLearner. Calibration involved subtracting the mean epigenetic age residuals of non-Nicoyans from those of Nicoyans. For comparison of the calibrated ensemble’s results, calibrations were also applied to five epigenetic clocks with the lowest MSE across the Nicoyan and non-Nicoyan populations: neural network centenarian 40 (NNCen40), DNAm fitness age (DNAmFitAge), PCGrimAge, DNAmGrimAge2, and ENCen40. NNCen40 and ENCen40 had been partially trained on about 45% of the CRELES data, giving them a potential performance advantage but also a risk of overfitting. The SuperLearner ensemble achieved the lowest MSE of 16.1, followed by NNCen40 at 25.2. Residual distribution plots revealed that only the SuperLearner ensemble indicated age deceleration among non-Nicoyans, whereas most other clocks indicated age acceleration among them. Furthermore, before calibration, the ensemble provided the strongest evidence of lower-than-expected epigenetic age in the Nicoyan population, estimating an epigenetic age approximately three years lower than chronological age. The NNCen40 clock also predicted biological age deceleration, with an effect about one-third as large, corresponding to a 1.17-year lower epigenetic age. By contrast, DNAmFitAge and PCGrimAge estimated a mean age acceleration in Nicoyans, whereas ENCen40 and DNAmGrimAge2 showed no significant differences. After calibration, the results were more consistent across clocks, with all suggesting younger epigenetic age in Nicoyans. Notably, the SuperLearner ensemble estimated the largest calibrated age deceleration relative to the reference population (1.96 years), followed by PCGrimAge (1.69 years), consistent with the SuperLearner assigning three-fourths of its weight to this algorithm. Next, the team adjusted Nicoyan's aging predictions for confounders that affect aging, such as body mass index (BMI), sex, socioeconomic status, and smoking. Confounder-adjusted results for each clock were similar to the calibration results. However, an additional age-stratified calibration substantially reduced the apparent Nicoyan advantage. Because exceptionally long-lived Nicoyans were oversampled, the researchers recalibrated estimates within 10-year age groups. The SuperLearner difference fell to just 0.35 years but remained statistically significant, while most fixed clocks no longer showed significant differences, suggesting that much of the apparent advantage was concentrated among the oldest participants and diminished when Nicoyans were compared with non-Nicoyans of similar ages. Conclusions Taken together, the team presented a novel SuperLearner framework for predicting epigenetic aging that integrates fixed epigenetic clocks with flexible ML algorithms and calibrates estimates against a matched reference population. Only the SuperLearner ensemble, trained on non-Nicoyans, suggested marginal age deceleration of approximately one year among non-Nicoyans. The uncalibrated analysis revealed limitations in the standard approach to epigenetic age estimation, with existing clocks estimating both age deceleration and acceleration in Nicoyans. However, calibrating Nicoyans to non-Nicoyans yielded consistent results, with all clocks estimating modest age deceleration of one to two years, with the SuperLearner ensemble producing the largest estimate. Adjusting for confounders affecting the aging process yielded an effect similar to that of calibration. When the researchers also accounted for differences in age distribution by comparing similarly aged participants, however, most of the apparent Nicoyan aging advantage disappeared, leaving only a small but statistically significant SuperLearner estimate of 0.35 years. As such, calibrating to a well-matched population may help reduce bias from unmeasured confounding, although survivor selection and age structure remain important considerations. *Important notice: medRxiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.