IntroductionAging is the measure of the gradual systemic degradation of the body’s physiological processes1. The degree of this decline may not be consistent with chronological aging, the temporal duration of the lifespan measured from birth. This is evidenced by the ability of lifestyle factors such as smoking, diet, and exercise to modulate health outcomes including diabetes2, cognitive decline3,4,5,6,7, and cancer in age-controlled studies. Recent efforts have been made to quantify age-related physiologic degradation and potential contributors of “accelerated” biological aging at the cellular level8. Prior studies attempting to quantify biological aging include the consideration of telomere length, inflammation, DNA damage and repair systems, mitochondrial activity/function, and epigenetic processes9. Of these, the “epigenetic clock” first introduced by Horvath in 201310 has garnered increased interest due to its widespread associations with morbidity and mortality; recent advancements have enhanced its accuracy by incorporating specific clinically relevant phenotypes and other biomarkers in newer generation clocks11,12,13,14.Epigenetic clocks were developed through machine learning on microarray-based DNA methylation (DNAm) data. The first-generation clocks, Horvath10 and Hannum11, accurately predicted age by analyzing the DNAm state at specific cytosines in a 5′-CG-3′ context, while second-generation clocks, such as Levine12 and GrimAge13,14, have further improved predictions related to aging and disease, by incorporating supplementary data into epigenetic age calculations. Considered the first of the third-generation clocks, DunedinPoAM15 and its successor DunedinPACE16, offer a novel perspective on epigenetic aging. They are the first epigenetic clocks to have been trained on a paired longitudinal cohort, allowing for the quantification of one’s ‘pace of biological aging’, rather than providing a simple biological age value. Despite the advances in these clocks, the mechanism(s) underlying the association between the DNAm states at clock-specific CpGs remain largely unknown.Prior studies of the intestinal (gut) microbiome have implicated it as a potential modifier and/or driver of biological aging17. Remarkably, fecal microbiota transplant (FMT) studies have reported decreases in age-acceleration following FMT to correct aging-associated dysbiosis in mice18,19. The gut microbiota is increasingly recognized as a critical determinant of health across a wide range of disease conditions. However, its compositional complexity and temporal variability present significant challenges for establishing robust associations between specific microbial taxa and health outcomes using conventional statistical approaches. The emergence of machine learning (ML) methodologies has enabled more sophisticated analyses, capable of capturing non-linear and high-dimensional relationships within the microbiome. Recent studies have demonstrated that ML models leveraging microbial relative abundance data can accurately predict various disease states20,21, highlighting their potential utility in microbiome-based diagnostics and risk stratification. One such application is the prediction of chronological aging (cAge). Microbiome data has demonstrated the ability to predict cAge within 5–10 years22.Given recent studies implicating the human gut microbiota as input in the prediction of disease states, and other studies describing the microbiome-epigenetic axis23,24, we tested the hypothesis that epigenetic aging may be estimated using gut microbiome data, which may serve as a valuable indicator of biological aging as its dynamic composition may already account for environmental factors, such as diet, that influences aging. To do so, we studied a cohort representing diverse ages (n = 123, 17–82 years old) where, importantly, stool and blood samples were collected from each participant at the same time point and from which we profiled the gut microbiome and genome-wide DNA methylation states within monocyte-enriched samples, respectively. Our results show for the first time that 16S-based (16S rRNA gene amplicon sequencing) metagenomic profiling data of the gut microbiome can estimate biological age measured from array-based DNAm of monocyte-enriched samples, strongly implicating the involvement of both novel and known microbial taxa in epigenetic aging.ResultsDemographics and epigenetic ageThis analysis includes a subset of data from adult participants recruited into the Hawaii Social Epigenomics of Early Diabetes (HI-SEED) study between 2021 and 2023 to better understand and address cardiometabolic health disparities. Each participant included here provided stool and blood samples on the same day at study entry, from which we examined gut microbiome composition using 16S-based metagenomic analysis (Figure S1) and monocyte-enriched blood cell DNA methylation using the Illumina EPIC array platform (Figure S1). We estimated epigenetic age (epiAge) using established epigenetic clocks and additionally applied microbiome data to predict epiAge using a de novo machine learning approach that we term “EpiBiome” for simplicity (Figure S2). Demographic data and conventional epiAge estimates are listed in Table 1. The cohort comprised 123 participants with a mean ± standard error of the mean (SEM) chronological age of 31.7 ± 1.5 years. Of these, 57 (46.3%) were male sex and 66 (53.7%) were female sex. Mean BMI was 31.2 ± 1.0 kg/m2: 41 (33.3%) participants were classified as normal weight, 24 (19.5%) as overweight, and 58 (47.2%) as obese according to current CDC thresholds. Mean HbA1c measured in blood using a point-of-care A1CNow® kit (PTS Diagnostics) was 5.6 ± 0.1%; 89 (72.4%) participants were non-diabetic, 20 (16.3%) were pre-diabetic, and 14 (11.4%) were type 2 diabetic based on current ADA cut points.Table 1 Study demographics.Full size tableAssociations between epigenetic age and common cardiometabolic risk factorsWe first characterized epigenetic aging metrics in this cohort and their relationships with metabolic health indicators (Fig. 1). The three traditional clocks (Horvath, Levine, GrimAge2) showed strong linear relationships with chronological age, as expected given their design as cross-sectional age estimators (Fig. 1a; Pearson r = 0.89, 0.82, and 0.90 respectively; all p