Using saliva bacteria and machine learning to estimate age and predict health without invasive tests

A schematic overview of the study design and the primary analytical approaches was provided in Figure 1A–E. Sequencing data were obtained from the National Center for Biotechnology Information (NCBI) Sequence Read Archive and databases from our team published previously (a total of 56 databases) focused on massively parallel sequencing of human 16S saliva samples (Table S1).

In total, 7268 samples were collected across six continents, including 493 infants, 1750 children, and 4385 adolescents/adults, ranging from 0 to 107 years (Figure, Table). 62.36% samples reported their health to be good or very good. Other samples were involved in several types of diseases and exposure statuses, such as viral infection, Obese related diseases, oral diseases or status (Figure). S1A S2 S1B

We utilized the QIIME2 pipeline for data import and preprocessing. All raw sequences were subjected to quality control using a consistent index and subsequently integrated into a unified data set.

To investigate the potential relationship between saliva microbiota and age, we calculated the alpha diversity of oral microbiota across different age groups. Shannon diversity indices increased across the age categories of 0, 1–4, and 5–9 years, followed by a stepwise decline in diversity in 15–24 years old, and another decrease in diversity was seen in over 65 years, indicating that saliva microbiota continues to develop until adolescent, experienced a small fluctuation in young adulthood, and declines in old age (Figure S1C,D). The relative abundances of the main genus changed with age. The percentage of Streptococcus was decreasing, while Prevotella and Veillonella were increasing with age (Table S3, Figure S1E,F). The above results revealed a potential association between saliva microbiota and age‐related changes.

We randomly split the healthy sample databases into 70% training and 30% test sets to develop a saliva MicroAge predicted model. During the training phase (n = 4532), we evaluated six machine‐learning methods to train saliva MicroAge predicted models for predicting chronological age using normalized expression composed of 8094 microbiome species (Figure S2). Among these methods, we found that gradient boosting (LightGBM) demonstrated the highest age prediction accuracy in the test set (R² = 0.7983, RMSE = 9.34, MAE = 5.92, Figure S2A, Table S4). The model also showed robust predictive performance across the four main continents in the test set (Figure S2B). Due to its superior generalizability, we selected LightGBM as our final model.

We also evaluated our model using an external data set from PRJNA1162741 [7]. Despite the significant differences in age distribution and no specific identifiers for healthy controls, our model yielded an intriguing result: we identified 136 individuals whose predicted MicroAge fell within ±5 years of their chronological age, a number close to the reported count of healthy controls in the original study (Figure S2C).

Next, we applied the saliva MicroAge prediction model to the whole databases, to investigate whether variations exist between salivary MicroAge and chronological age in varying disease conditions and extreme environments. The results revealed a lower correlation between MicroAge and chronological age in unhealthy samples (R² = 0.1642, RMSE = 21.16, MAE = 16.16; Figure S2D), compared to healthy populations.

This trend was consistently observed across specific disease groups, including Head and neck cancer, Schizophrenia, Asthma, and COVID‐19 infection samples (Figure), suggesting that the model captures variance primarily driven by disease‐related factors. Other machine‐learning models are shown in Figure. S2E S3

To understand how saliva microbial aging may influence human diseases and various environmental exposure states, we categorized disease types and exposure environments in the database samples and investigated the association between predicted saliva MicroAge and these states (Figure 2A). Mean squared error (MSE) was applied to evaluate the generalization ability of MicroAge model compared to chronological age under different states. We found that MicroAge exhibited the least fluctuation under healthy states, whereas its variability increased under disease states, suggesting that the model performs most consistently under healthy states. Among all states, the fluctuation of MicroAge was the largest under viral infection, excluding residuals state (Figure 2B). Detailed effect size (Cohen's d) results are provided in Table S5. The volatility of saliva MicroAge compared to chronological age may be a powerful predictor of health states.

To further investigate, we defined saliva MicroAgeGap as the difference between saliva MicroAge and chronological age, and further used multiple linear regression models to investigate whether associations of MicroAgeGap with different health states included in our database (Figure 2C). Healthy state was set as the reference benchmark. The forest plots were divided into three sections by two dashed lines according to regression coefficient. The states in the upper section have positive regression coefficients, indicating a positive effect on predicted age. These states were associated with inflammation, such as rampant caries, ulcerative colitis, and adenocarcinoma. The middle section represents conditions with minimal impact on predicted age, including diseases such as dental caries and oral submucosal fibrosis. The lower section contains diseases with negative regression coefficients, indicating a negative effect on predicted age. Many of these were related to immune deficiencies, such as HIV and diffuse large B‐cell lymphoma.

Our findings suggested that a positive increase in saliva MicroAgeGap relative to the healthy baseline may be linked to aging in inflammatory states, whereas a negative increase may indicate a decline in immune function. MicroAgeGap might be a powerful predictor of health states.

Then, we explored the association between MicroAgeGap and oral diseases and focused on two important oral diseases: dental caries and periodontitis [8, 9, 10].

For dental caries, the fluctuation of MicroAgeGap in common dental caries and rampant caries was the smallest, while the fluctuation of arrested caries and rampant caries accompanied by high‐sugar diet was the largest (Figure). S4A

For periodontitis, severe periodontitis with HIV infection had the highest fluctuation of MicroAgeGap, and there was a statistically significant difference in MicroAgeGap between the severe periodontitis and healthy groups (Figure). In addition, we found that as the severity of periodontitis increased, MicroAgeGap also gradually increased (Figure). S4B S4C

Subsequently, we investigated the connection between MicroAge and exceptional longevity. We collected and downloaded 64 saliva samples from the PRJEB25916 data set (Figure S5A), which included a younger, elder age, and centenarian groups (Figure S5B–D) [11]. Centenarian group had a larger MicroAgeGap compared to both the younger and elder age groups, meaning that their predicted age was considerably younger than their actual age. The results suggest that centenarians have a protective MicroAge phenotype, indicating that MicroAge is associated with exceptional longevity and might provide a more accurate indication of true biological age compared to chronological age [12].

To determine which microbial species were most critical for predicting chronological age among all microbial species included, we identified the top 30 most important features in the LightGBM model. The most important feature was Parvimonas micra, an oral opportunistic pathogen frequently associated with oral diseases such as periodontitis, apical periodontitis [13, 14, 15]. Notably, some species with relatively low abundances in saliva, such as Porphyromonas gingivalis, were also among the top 30 important features and closely linked to diseases like periodontitis [16, 17], highlighting the clinical relevance of our model (Figure S6A).

Then, we used the top 30 features to construct a prediction model, and the model demonstrated robust predictive performance (Figure). Furthermore, when applying the top 30 features in a multiple linear regression model to predict MicroAgeGap in relation to diseases and environmental exposures, the simplified model achieved a correlation coefficient similar to the full‐feature model, further validating its utility (Figure). In the microbial networks, we observed that top 30 important features consisently linked with other features across four continents (Figure). S6B S6C S6D

We predicted the potential functions of the top 30 important features using Tax4Fun (Figure). In the 0–9 age group, the functions of the top 30 important features were associated with terms such as metabolism, genetic information processing. In the middle‐aged and elderly groups, the functions of the top 30 important features were linked to aging, oxidative phosphorylation, and immune diseases. Our findings suggest that the core salivary microbiota might have a role in co‐developing alongside human biological processes. S7

Our study analyzed 7268 16S salivary microbiome samples from six continents, assessing microbial diversity and abundance across different age groups to explore their relationship with age. We then developed and validated a machine learning model to predict saliva MicroAge, demonstrating strong predictive accuracy in healthy populations. We also investigated the variability between MicroAge and chronological age (MicroAgeGap), revealing its association with various health states.

Our approach offers several advantages, including the use of machine learning models and microbiome data, which effectively capture nonlinear relationships and interactions within the oral microbiome. Additionally, our model evaluation demonstrates that our microbial age prediction model significantly outperforms traditional methods, providing enhanced accuracy and generalizability in predicting biological age compared to other algorithms like LASSO and elastic net [2].

However, there were still some limitations in our study, and future research directions to investigate. First, the sample size remained relatively small. In future research, we aimed to expand our database to develop a more robust model. Second, our data were from the NCBI database. Several specific sample information, such as saliva sample storage, time, environmental factors and preprocessing, antibiotic use, and social relationships, remained unknown. In future studies, we will integrate clinical research and experimental approaches to explore the mechanistic explanation in greater depth. Large language models will be applied in our prediction model due to their impressive capabilities.