Human Skin, Oral, and Gut Microbiomes Predict Chronological Age

Considerable evidence suggests that the gut microbiome changes with age or even accelerates aging in adults. Whether the age-related changes in the gut microbiome are more or less prominent than those for other body sites and whether predictions can be made about a person’s age from a microbiome sample remain unknown. We therefore combined several large studies from different countries to determine which body site’s microbiome could most accurately predict age. We found that the skin was the best, on average yielding predictions within 4 years of chronological age. This study sets the stage for future research on the role of the microbiome in accelerating or decelerating the aging process and in the susceptibility for age-related diseases.

KEYWORDS age prediction, gut microbiota, oral microbiota, random forests, skin microbiota M icrobiomes across the body are known to change rapidly in the first 3 years of life and then relatively little in adults (1). Recent work suggests that the gut microbiome can be used to classify adults into age groups, with sex-specific differences in patterns of diversity by age (2). Age has been implicated as a dominant factor in the adult microbiome in numerous cohort studies (3)(4)(5). The microbiome also continues to change after death and has been used to predict postmortem interval to within a few days in both mice (6) and humans (7): contrary to expectation, the skin microbiome predicted postmortem interval much better than did the gut or the surrounding soil microbiome.
We acquired 100-bp amplicon sequence variants (ASVs) processed with Deblur (15) from the 16S-V4 rRNA gene amplicon data in Qiita (16) using the redbiom search engine (17). This study includes only subjects with self-reported ages from 18 to 90 years (see Fig. S1 in the supplemental material), body mass indices (BMI) of 18.5 to 30 kg/m 2 , no reported inflammatory bowel disease or diabetes, and no antibiotic consumption 1 month before sampling. We also excluded pregnant, hospitalized, disabled, or critically ill individuals (Table S1). For gut microbiota, the majority of acquired samples were derived from two projects: (i) the American Gut Project (AGP) (3) and (ii) the Guangdong Gut Microbiome Project (GGMP) (13). For oral and skin microbiota, we obtained all samples from Qiita matching the inclusion and exclusion criteria above, representing the most comprehensive meta-analysis for age prediction using human microbiota. We further analyzed the ASV data with the QIIME 2 pipeline (18).
We used random forests (RF) (19) to regress relative abundances of ASVs in the healthy human microbiota from different body sites (gut, oral, and skin microbiota) against the subjects' chronological ages with the R package ranger (20) using finetuned hyperparameters. To test if confounders (such as sex) affected the modeling, we first trained the age model within a sub-data set stratified by a confounder and then applied it on all the other sub-data sets. For both model training and testing, we evaluated regression performance using mean absolute error (MAE). We fit a smoothing spline function between microbiota age and chronological age to calculate relative microbiota age. Relative microbiota age per sample was calculated as the difference between the microbiota age of a focal adult and the microbiota age of the interpolated spline fit of healthy adults at the same chronological age. We used the Wilcoxon rank sum test (21) to compare relative microbiota ages between host groups in each data set. To determine the effects of country and body sites on microbiota age, we subdivided the data sets into these groups and repeated the analyses.
The RF regression recaptured the known result that the gut microbiome is associated with chronological age (Fig. 1A) and that this relationship holds across cohorts, but the connections to age were even stronger in the oral (Fig. 1B) and skin (Fig. 1C) microbiomes. Remarkably, the skin microbiome could pinpoint a subject's age to within 4 years, on average. Analysis of the specific microbial features contributing to these models demonstrated that relatively few ASVs (e.g., around 64) are needed for highly accurate models for each body site (Fig. 1D to F).
We next tested whether the models were sex specific. As shown in previous work (2), we found a sex-specific signal in the gut microbiome; however, we did not find a sex-specific signal in the mouth or the skin microbiome. Consequently, although we observed a small degradation in prediction accuracy for the cross-trained models from men to women or vice versa for the gut, we saw no such degradation for the other body sites, suggesting that populations do not need to be stratified by sex to build such predictive models.
For the skin, we had enough data for the forehead and palm to test whether models trained on skin from one body site apply for the other. This is important because the forehead and hand are markedly different in physiology and microbiology (22). Figure 2A and B demonstrate that models of microbiome age for the forehead can be cross-trained on the palm, and the converse is also true (Fig. 2C). This means that future studies seeking to determine factors leading to microbiome aging can combine these skin sites, which is important given the tremendous microbiome and metabolic diversity observed across the human body (23).
An important consideration is which taxa contribute to the age prediction model. In the gut, the ASVs belonging to the genera Bifidobacterium and Blautia or the families Lachnospiraceae, Ruminococcaceae, and Clostridiaceae consistently had high feature importance scores, although values differed between populations and within populations (Fig. S2). A larger discrepancy in feature importance rankings was found between aging models built from different countries. For example, the top-ranking feature in gut microbiota MAE: 11.5 ± 0.12 yr R squared: 16  Chinese cohorts is an ASV belonging to Bifidobacterium, but it was not detectable in the U.S. cohort. In the oral microbiota, we identified a set of top-ranking microbial markers decreasing in abundance with host aging in both females and males, such as ASVs belonging to Lactobacillales, Gemellaceae, Bacteroides, and Fusobacterium (Fig. S3). In the skin microbiome, we identified age-related markers in four subgroups: female forehead, male forehead, female palm, and male palm. As we age, changes in skin physiology (such as decreased sebum production and increased dryness) and host immune system can alter associated microbiota (24)(25)(26). Interestingly, we identified several genera and families that include anaerobic members (i.e., ASVs belonging to Mycoplasma, Enterobacteriaceae, and Pasteurellaceae) negatively correlated with age in all subgroups, reflecting these physiological changes due to aging (Fig. S4). These results are consistent with city-specific influences on models for predicting clinical states (13). However, the success of the cross-population generalization suggests that the types of tuned RF models we introduce here may result in robust and generalizable predictors. Interestingly, gut and oral bacteria that are enriched in young individuals are both more abundant and more prevalent than bacteria enriched in elderly individuals. We calculated the average relative abundance and ubiquity for each ASV in the shared response. We found that bacteria that are enriched in young individuals in at least two cohorts are both ubiquitous and abundant across people, whereas those enriched in old age are less abundant and not ubiquitous. Thus, the presence of these ASVs enriched in elderly individuals is a good indicator of microbial shifts associated with aging.
Taken together, our results demonstrate that accurate and generalizable indicators of age can be derived from microbiome studies using machine learning techniques and that prediction is most accurate from the skin microbiome. Building on these results, future work will include developing noninvasive microbiome-based tests to determine signs of accelerated or delayed aging in the elderly, or in individuals with chronic diseases, and designing and evaluating microbially based interventions to modify the aging process.
Data availability. The data and code for this study are available at https://github .com/shihuang047/age-prediction.

SUPPLEMENTAL MATERIAL
Supplemental material is available online only. FIG S1, PDF