Statistical Shield Gives Epigenetic Age Clocks Honest Error Bars

Epigenetic clocks can read a person’s approximate age from the pattern of chemical tags on their DNA, but most of these tools hand over a single number with no honest statement of how far off it might be. A new study in GeroScience addresses that gap by wrapping a clock called BayesAge in a statistical framework known as conformal prediction, which produces prediction intervals with mathematically guaranteed coverage under stated assumptions. The work, led by Megan Mitchell, Lajoyce Mboning, Louis-S. Bouchard and Matteo Pellegrini, shows that a relatively simple calibration step can turn point estimates from a methylation-based age predictor into intervals that contain the true age roughly as often as promised.

The underlying predictor, BayesAge, is a maximum-likelihood method that the same group previously introduced. Unlike the familiar epigenetic clocks built by Horvath and Hannum, which use penalized linear regression on panels of 353 and 71 CpG sites respectively, BayesAge models the relationship between DNA methylation and age as nonlinear. It fits locally weighted scatterplot smoothing curves, known as LOWESS, relating methylation fraction to age at each selected locus, and then treats methylated read counts at those loci as binomial count data. The method scans a discrete grid of candidate ages from zero to 100 and picks the age whose likelihood best explains the observed methylation counts, using only 16 CpG sites in the configuration tested here.

Conformal prediction works by holding out a calibration set of labeled samples whose true ages are known. The predictor is fit on a training subset, and the absolute errors between predicted and true ages are computed for every calibration sample. These residuals are sorted, and a quantile is chosen so that, under the assumption that calibration and test samples are exchangeable, a new prediction interval will contain the true age at least a target fraction of the time. Crucially, this guarantee is distribution-free: it does not require assuming any particular shape for the error distribution, and it holds with finite samples rather than only in the limit. Split conformal prediction, the variant used here, needs just one model fit per data split, making it computationally cheap compared with full conformal approaches that refit the model for every candidate label.

To test the approach, the researchers assembled targeted bisulfite sequencing data from 988 subjects, aggregated from multiple studies with a shared probe set enriched for age-associated CpG sites. Specimens included blood, buccal swabs and saliva, and the processed data yielded a matrix of 10,710 CpG sites across all samples. They ran 12 cross-validation splits in which 200 samples per split served as the calibration set, the remainder was divided between training and validation, and each sample appeared as a validation point exactly once. At a target coverage level of 0.9, split conformalized BayesAge achieved empirical coverage of 91.2 percent, close to the nominal 90 percent, while the point-prediction accuracy barely moved: the mean absolute error was 6.083 years with conformalization versus 6.098 years without it, despite the calibration set shrinking the effective training data.

Before bounding intervals to the admissible age range of zero to 100 years, the average half-width of the conformal intervals was 12.984 years, meaning a typical reported interval spanned roughly 26 years of age. After truncation to the valid range, the mean total width was 25.61 years. The authors note that bounding can make intervals asymmetric near the boundaries, for example when a predicted age sits close to zero, but it does not change whether the interval contains the true age when that age lies within the admissible range. A small-sample stress test reinforced the theoretical claim: with only 10 training, 10 calibration and 10 validation samples per split, repeated over 1000 random splits, the running average empirical coverage stabilized near the 0.9 target, consistent with the finite-sample guarantee.

Constant-width intervals, however, can hide an important feature of methylation-based age prediction: errors may grow with age. To capture this heteroskedasticity, the team also evaluated a locally scaled variant in which each residual is divided by a locally estimated error scale, obtained by smoothing the relationship between predicted age and absolute error with LOWESS. The resulting intervals widen at older predicted ages, and in the main experiments the locally scaled method reached 89.6 percent empirical coverage with an average bounded width of 24.868 years. Yet this flexibility comes at a cost. Because the scale function itself is estimated from the calibration residuals, the standard coverage guarantee no longer strictly applies, and in the small-sample stress test the locally scaled method’s empirical coverage failed to approach the nominal level, revealing weaker coverage control when calibration sets are tiny.

The study also compared conformalized BayesAge against three alternative ways of attaching uncertainty to age predictions. The first was the Monte Carlo simulation procedure from the original BayesAge implementation, which generates synthetic methylation counts at each locus by binomial sampling and propagates them through the predictor to obtain a distribution of simulated ages. In this dataset that approach under-covered badly, achieving only 42.3 percent empirical coverage at a nominal 90 percent. The authors attribute this to the fact that the simulation primarily propagates read-sampling variability conditional on the observed methylation fractions, while ignoring model error, cohort heterogeneity and other uncertainty sources that dominate in practice.

The second and third baselines were higher-dimensional linear models wrapped in the same conformal machinery: conformalized linear quantile regression using 88 CpG sites, and split conformalized Lasso regression with a penalty tuned to 0.02, whose fitted models used an average of 244.5 CpG sites per split. The quantile regression baseline achieved 90.3 percent coverage with a mean bounded width of 23.18 years, while conformalized Lasso produced the narrowest intervals of all, 16.33 years on average, with 90.2 percent coverage. These linear baselines also delivered lower point-prediction error in this dataset, with mean absolute errors of 4.694 years for quantile regression and 3.925 years for Lasso, compared with 6.083 years for conformalized BayesAge. The trade-off, the authors argue, is that BayesAge reaches calibrated uncertainty with substantially fewer CpG sites, which reduces assay complexity and computational cost.

Residual structure added another dimension to the comparison. Using signed residuals, defined as predicted age minus chronological age, the team found that BayesAge showed less age-dependent structure than the linear baselines, which displayed increasingly negative residuals at older ages. Because conformal prediction wraps a point predictor without altering its point estimates, these patterns reflect differences among the underlying models rather than any effect of conformalization. The authors suggest that residual behavior deserves attention alongside mean absolute error and interval width, since many applications focus on age acceleration, the difference between predicted and chronological age, and age-dependent residual bias can complicate its interpretation. They also caution that residual structure is dataset- and implementation-dependent and should be reassessed whenever a clock is deployed in a new cohort.

The researchers are explicit about the limitations. The dataset was aggregated from projects with diverse objectives, specimen types and potential batch effects, and such heterogeneity can challenge the exchangeability assumption on which the coverage guarantee rests, so coverage should be checked within relevant subgroups such as tissue, study or age range. Several design choices, including the number of CpG sites and the LOWESS smoothing parameter for the local scale, were selected through exploratory analyses in the same cohort, and a nested selection strategy or an external hold-out cohort would provide cleaner separation between model selection and evaluation. The guarantee is also marginal over calibration draws and test examples, not conditional on covariates, meaning calibrated intervals may still under-cover within specific strata. Even so, the authors conclude that split conformal prediction offers a computationally simple, distribution-free route to uncertainty quantification that can wrap any point predictor, including likelihood-based clocks capturing nonlinear methylation-age relationships, and could support applications where quantified error bounds are needed to interpret age estimates and age acceleration measures.

Subject of Research: Calibrated uncertainty quantification for epigenetic age prediction using conformal prediction with BayesAge

Article Title: Conformal uncertainty quantification for BayesAge epigenetic age predictions

Article References: Mitchell, M., Mboning, L., Bouchard, L.-S., & Pellegrini, M. (2026). Conformal uncertainty quantification for BayesAge epigenetic age predictions. GeroScience. https://doi.org/10.1007/s11357-026-02600-5

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02600-5

Keywords: epigenetic clocks, BayesAge, conformal prediction, DNA methylation, prediction intervals, uncertainty quantification, aging biomarkers, bisulfite sequencing, GeroScience, age prediction, coverage guarantees, Conformal