Synergistic and heterogeneous aging using composite phenotypes and multiple organ systems aging clocks

Abstract

Background:

Assessing aging pace through biological age offers a precise perspective and underscores the need for further investigation into organ-level disparities.

Methods

This observational study utilized multi-scale phenotypes from the Taizhou Imaging Study, encompassing brain imaging, cognitive assessment, blood biochemistry, omics, and physical measures. A total of 904 individuals (403 men and 501 women) aged 55-65 years were included. Age correlations with single and composite phenotypes were assessed, and multi-modal aging clocks were developed, incorporating organ systems, cognition, and the whole body.

Results

Here we show that composite phenotypes, such as those of cardiovascular system and bone, alter with age progression and could serve as aging clock features. Despite existing connections among various organs’ aging rates, their low intensity (under 0.25) indicates the variability of aging. Accelerated aging in the brain (mediating 12.46%, 95% CI: 4.37% to 24.44%) and kidneys (mediating 6.94%, 95% CI: 1.08% to 18.63%) partially mediates the relationship between smoking and the decline in olfactory identification. The diversity of organ aging is also evident as accelerated aging extends from the cardiovascular system to the kidneys and brain with increasing metabolic risk factors. Moreover, the biological ages of cardiovascular system, bone, metabolism, brain and the whole body show stronger associations with cardiovascular events risk than chronological age.

Conclusions

The overlap between composite phenotypes and biological age provides valuable insights into multi-phenotypic aging. The unearthing of the heterogeneity in aging processes could further inform the development of personalized interventions to slow organ-specific aging and better manage age-related health problems.

Plain language summary

Aging affects different parts of the body in different ways, but most studies focus on single organs or single health measures. This study aimed to understand how aging unfolds across multiple organs and systems, and how these patterns relate to lifestyle factors and future disease risk. Using health and imaging data from middle-aged adults in rural China, we built “aging clocks” that track changes in traits across organs such as the heart, brain, bones, and metabolism. We found that some organs age faster than expected for a person’s age, and this uneven aging was linked to lifestyle risks like smoking and to future heart disease. These findings highlight the importance of targeting organ-specific aging to improve health in later life.

Li et al. analyze multi-scale phenotypes to investigate heterogeneity in multi-organ aging patterns. They find that accelerated aging in organs such as the cardiovascular system, brain, kidneys, bone, and metabolism is associated with organ-specific links to lifestyle factors, health status, and cardiovascular risk.

Introduction

Age is the main risk factor for diseases such as neurodegeneration and cardiovascular diseases¹. An accelerated or premature aging process is intimately linked to health conditions across an individual’s lifespan², highlighting the necessity for precise aging assessment to identify potential health risks and initiate early interventions. Biological age, an effective tool for quantifying aging, allows for a more authentic evaluation of the pace of aging compared to chronological age³. Estimation models for biological age, also known as aging clocks, have been constructed considering a variety of factors beyond chronological age, including omics^4,5, clinical phenotypes^3,6–11, and other biomarkers. These models capture the cumulative effects of aging processes and highlight inter-individual differences in aging among peers.

Existing aging clocks primarily assess the rate of aging at an individual level^{3,6,7,12–14}. However, animal research suggest that aging not only varies among individuals but also encompasses different molecular aging trajectories among various organs within the same organism^15,16. Emerging research indicate that such organ-specific disparities also exist in humans^4,11,17. For instance, Oh et al.employed plasma proteomics to estimate aging in 11 organs, revealing distinct associations between accelerated organ aging and Alzheimer’s disease progression⁵. While some studies have focused on individual organs, such comprehensive investigation into organ aging heterogeneity within the same population is needed to better capture the variability in aging patterns. Furthermore, identifying modifiable factors and their disease links is crucial to fully unlocking the clinical potential of aging clocks. Emerging deep phenotyping^18–20 technologies have expedited the collection of detailed individual data. Composite phenotype²¹, crucial in data interpretation, identifies interrelated traits relevant to a specific condition. In aging research, the composite phenotype is promising for unveiling potential biological markers, identified as a mix of aging-related traits²². Despite this, current studies remain inadequate.

Here, our study is anchored on the multi-scale phenotypes derived from a rural late middle-aged population in China, excavating the synergistic and heterogeneous aging characteristics masked beneath, as well as identifying the related factors. Based on the Taizhou Imaging Study (TIS)²³, a community-based prospective cohort, the objective of this study specifically lies in examining the following aspects within the sphere of biological aging: (1) the variability of individual and composite phenotypes with progressing age; (2) the interconnections and distributional discrepancies between multi-modal aging patterns at the three levels: organ systems, cognition, and the whole body; (3) several exposure factors and health status associated with distinct aging clocks; (4) the practical application of the aging clocks in predicting the risk of future cardiovascular events. We show that both single and composite phenotypic measures capture age-related changes and can serve as features of aging clocks. Organ-specific biological ages reveal synchronous yet heterogeneous aging patterns, with accelerated brain and kidney aging mediating functional decline and cardiovascular, bone, metabolism, brain and whole-body ages showing stronger associations with cardiovascular risk than chronological age. Together, these findings establish the potential of multi-organ aging clocks for capturing the variability of aging, linking modifiable exposures with organ-specific trajectories, and improving disease risk prediction.

Methods

Data source

This observational study utilized the baseline and the second-round follow-up from the TIS²³. The baseline examination was initiated from 2013 to 2018 in China, Taixing. At baseline, the TIS enrolled participants aged 55-65 years who had lived in Taizhou for at least the last 10 years, and had not been diagnosed with dementia, stroke, cancer, psychiatric disorders, severe liver disease, or kidney disease. Those with contraindications to MRI or known claustrophobia were also excluded. Additionally, individuals had to be able to walk, communicate, and provide information independently. Finally, a total of 904 individuals, including 403 men and 501 women, were included in the baseline of TIS. After a median six-year follow-up, 200 participants, for reasons including death post-baseline survey, severe illness, refusal, migrant work commitments, and lost contact, failed to undergo the second follow-up examination. Consequently, 704 participants were included in the longitudinal analyses, with 304 men and 400 women.

This study received approval by the Ethics Committee of the School of Life Sciences, Fudan University, and Fudan University Taizhou Institute of Health Sciences (institutional review board approval number: 496 and B017, respectively). The participants provided their written informed consent to participate in this study.

Phenotypes employed in the baseline

A set of phenotypes of all the baseline participants were employed, across 12 categories: brain MRI, cognitive assessment, blood biochemistry, two categories within the omics (gut microbiome, metabolome), and seven categories of physical measures, encompassing bone mineral density (BMD), anthropometry, olfactory, gait, blood pressure, arterial ultrasonic, and electrocardiography. To obtain a consistent and large dataset, maximizing the power of subsequent correlation calculations, we handled missing data within each category separately. Within each category, features with a missing rate exceeding 20% were first removed, followed by removing samples with a missing rate exceeding 20%. Only the average values were retained when phenotypes from both the left and right sides of the body were assessed, such as ankle-brachial index (ABI), brachial-ankle pulse wave velocity (baPWV), and systolic or diastolic blood pressure (SBP or DBP). Yet this averaging procedure was not applied to brain MRI phenotypes. Subsequently, within each category, multivariate single imputation using the k-nearest neighbors (KNN) method was performed. All variables underwent Z-score standardization, except for the gut microbiota, which was processed with DESeq normalization²⁴. After processing, a total of 514 phenotypes across 12 categories were included in the final analysis. The number of phenotypes of each category, along with their corresponding sample sizes, are as follows: brain MRI (153 phenotypes, n = 799, 346 men), cognitive assessment (8 phenotypes, n = 786, 350 men), blood biochemistry (12 phenotypes, n = 844, 370 men), BMD (13 phenotypes, n = 855, 382 men), the gut microbiome (123 phenotypes, n = 656, 290 men), metabolome (178 phenotypes, n = 869, 382 men), anthropometry (6 phenotypes, n = 826, 358 men), olfactory (1 phenotype, n = 904, 306 men), gait (2 phenotypes, n = 721, 306 men), blood pressure (10 phenotypes, n = 838, 370 men), arterial ultrasonic (2 phenotypes, n = 869, 388 men), and electrocardiography (6 phenotypes, n = 904, 403 men) (Supplementary Data 1).

MRI protocol and imaging analysis

All brain MRI scans were conducted on a single 3.0-T Siemens Magnetom Verio Tim scanner with a predetermined protocol, consisting of (1) a localizer scan; (2) susceptibility-weighted imaging (SWI) with a voxel size of 0.7 mm × 0.7 mm × 1.5 mm; (3) three-dimensional T1 magnetization-prepared rapid gradient-echo (MP-RAGE) with a voxel size of 1.0 mm × 1.0 mm × 1.0 mm; and (4) fluid-attenuated inversion recovery (FLAIR) with a voxel size of 0.9 mm × 0.9 mm × 3.0 mm. Detailed information on the acquisition parameters can be found in Supplementary Table 1. Structural T1-weighted scans were obtained for every participant at each site.

The Desikan-Killiany cortical atlas was employed to derive the features²⁵. Using FreeSurfer 7.2.0 software, 68 measures of cortical thickness, 68 measures of surface area, 14 subcortical gray matter regions (including the putamen, amygdala, caudate, nucleus accumbens, hippocampus, pallidum, and thalamus), 2 lateral ventricles, and estimated total intracranial volume were segmented. Note that all these measurements were conducted for both the left and right sides.

Cognitive function assessment

The assessment of cognitive function was conducted through the utilization of the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) questionnaires, both of which were employed in our previous studies^26,27. The MMSE total score was used as a summary measure. The MoCA was analyzed using both the total score and six cognitive domains, derived by grouping its 30 items into six distinct cognitive domains: short-term memory, attention/concentration/working memory, language, visuospatial ability, executive function, and orientation²⁸. Details of the cognitive domains and tasks can be found in Supplementary Table 2.

Blood biochemistry

Blood samples, obtained after an overnight fast (over 8 h), were collected by licensed nurses. Subsequently, clinical lab-based tests were conducted to measure serum parameters, including alanine aminotransferase, direct bilirubin, creatinine, urea nitrogen, uric acid, glucose, total cholesterol, triglycerides, high-density lipoprotein (HDL), low-density lipoprotein (LDL), total bilirubin, and cystatin C.

Metagenomic sequencing and profiling

The sample collection, metagenomic sequencing, and profiling process have been previously reported^23,29. In brief, sequencing was conducted on a NovaSeq platform (Illumina, USA) in a 150-pair-end format, resulting in an average sequencing depth of 3.6 gigabytes or approximately 24.5 million reads per sample. After quality control with KneadData version 0.10.0, the clean reads were mapped to the ChocoPhlAn pangenome database using MetaPhlan v3.0.13 with default parameters, yielding a taxonomic profile. After removing taxa with a prevalence of less than 10%, the remaining taxonomic data included 123 taxa at the species level.

Metabolomic profiling

Metabolomic profiling was conducted using a 600-MHz nuclear magnetic resonance (NMR) spectrometer equipped with an automatic sample jet (Bruker Biospin, Germany) at the metabonomic platform of the Human Phenome Institute, Fudan University²³. For each sample, we obtained three distinct one-dimensional spectra using the NOESYPR1D sequence, Carr-Purcell-Meiboom-Gill sequence, and a diffusion-edited spectrum. Additionally, we gathered a series of two-dimensional spectra from pooled samples to aid in spectral assignment. The high-throughput NMR metabolomic platform facilitated spectra generation, enabling the simultaneous quantification of approximately 350 metabolic features and ratio parameters. These were quantified either as absolute concentrations for each feature or as calculated ratios, encompassing lipids, lipoproteins, amino acids, and various other metabolites, but only absolute quantitative indicators were used for this analysis. All the analyzed serums were not subject to a second freeze–thaw cycle before metabolomic analysis. The detailed annotations of the metabolite names are demonstrated in Supplementary Data 2.

Bone mineral density assessment

A dual-energy X-ray absorptiometry, the Lunar DPX NT-400157 model manufactured by GE Healthcare in Madison, WI, United States, was employed to measure BMD. The assessed regions included the individual lumbar spine from the first to the fourth vertebra (L1–L4), the central portion of a lateral scout view of L1 to L4, femoral neck (with lower and upper segments specified as neck.low and neck.up, respectively), femoral shaft, trochanteric region, and Ward’s triangle, as well as the full hip. The measurements were conducted strictly following the manufacturer’s instructions, and the same experienced physician used the same machine for evaluation.

Other phenotypes of physical measures

Anthropometry, olfactory function assessment, motor function assessment, blood pressure measurement, arterial ultrasonic, and electrocardiography were also conducted by experienced professionals, including technicians, physicians, or sonographers²³. Anthropometry includes height, weight, waist circumference, hip circumference, body mass index (calculated as weight(kg)/height(m)²), and waist-hip ratio (calculated as waist circumference/hip circumference). The olfactory function of participants was assessed using the Connecticut Chemosensory Clinical Research Center (CCCRC) olfactory test, which includes a butanol odor threshold test and an odor identification test with seven common odors, with olfactory identification scores utilized in this study. The blood pressure measurements were obtained from two sources. The first involved traditional cuff and sphygmomanometer instruments (HEM-7210, OMRON HEALTHCARE Co., Ltd., Tokyo, Japan), with two automated measures taken 3–5 min apart. The second source employed ankle-brachial blood pressure measurements, utilizing a color Doppler B-mode ultrasound diagnostic scanner (Acuson S2000; Siemens AG, Munich, Germany) with automatic waveform analysis. Additionally, ABI and baPWV were determined using the same ultrasound diagnostic scanner. The instrument used for 12-lead electrocardiography measurements was the ECG-1350P by Nihon Kohden, Japan. After excluding parameters with more than 20% missing data, this study focused on the following indicators: heart rate; QRS duration; QT duration; RV5; SV1; and the PR interval. The assessment of gait function involved in this study included two gait tests: the timed up and go (TUG) test and the Tinetti test, both previously reported in this cohort³⁰. The TUG measures basic mobility by timing how long it takes to stand up from a chair, walk 3 meters, turn, and return³¹. The Tinetti test assesses balance and gait through a 28-point scale, with lower scores indicate poorer performance³².

Generation of composite phenotypes

To extract composite phenotypes, we performed a series of steps: constructing a phenotype association network, detecting modules, evaluating their robustness, and finally annotating the resulting modules with biological information. Pairwise similarities were calculated using partial correlations adjusted for age and sex. The resulting similarity matrix was transformed into a connectivity strength matrix using the soft-thresholding method in weighted correlation network analysis³³, producing a weighted association network of phenotypes. The soft-thresholding power was selected to ensure the resulting network conformed to a scale-free topology, a common characteristic of biological networks. We evaluated a range of powers from 1 to 20 by plotting two diagnostic metrics: the scale-free topology model fit and mean connectivity. A power of 4 was selected to balance high scale-free fit with the preservation of relevant phenotypic associations and network topology (Supplementary Fig. 1).

To segment the phenotype association network, we applied the Louvain algorithm, a classical approach for community detection in biological complex networks³⁴. The network was partitioned 101 times using this method, with the partition yielding the highest modularity selected as the benchmark. Additional community detection algorithms, including Leiden, Infomap, and Walktrap (implemented via the igraph R package), were also applied 100 times each for comparison. The resulting partitions were evaluated based on structural metrics, including modularity and conductance. Modularity quantifies the strength of a network’s division into modules, with higher values indicating more distinct separation. Conductance measures the proportion of edge weights connecting a module to the rest of the network relative to the module’s total edge weights, with lower values reflecting stronger internal cohesion and greater isolation from other modules. The Louvain algorithm achieved an optimal balance of high modularity and low conductance, while producing interpretable community sizes, supporting its selection as the primary method for module identification (Supplementary Fig. 2).

To further evaluate the robustness of the modules identified by the Louvain benchmark partition, we assessed their similarity to modules derived from the alternative algorithms. The similarity between modules was computed using a similarity matrix, where rows represent modules from the benchmark partition and columns represent modules from the compared partitions. Matrix entries symbolize the degree of similarity between the modules. Specifically, for different partition results, the calculation of the module similarity matrix was as follows:

Suppose we have two partitions $M$ and $P$, where $M$ has $m$ modules and $P$ has $p$ modules. We can define an $m\times p$ similarity matrix $S$, where the element $s_{ij}$ represents the similarity between the $i$-th module in $M$ and the $j$-th module in $P$. The calculation formula for $s_{ij}$ is as follows:

$$s_{ij}=\left\{\begin{array}{cc}\hfill 1,\hfill & \hfill \text{if}∣M_i\Delta P_j∣=0\hfill \\ \hfill \frac{∣M_i\cap P_j∣}{∣M_i\Delta P_j∣},\hfill & \hfill \text{otherwise}\hfill \end{array}\right)$$

where $M_i$ and $P_j$ are modules in $M$ and $P$ respectively, $∣M_i\cap P_j∣$ is the number of elements in the intersection of module $M_i$ and $P_j$, and $∣M_i\Delta P_j∣$ is the number of elements in the symmetric difference of module $M_i$ and $P_j$.

The maximum value in each row of the similarity matrix was selected to represent the maximum similarity between each module in the benchmark partition and the corresponding modules in the compared partitions. A larger value indicated a higher similarity between the module in the benchmark partition and the modules generated by other partition methods, signifying better robustness of that module. For each benchmark module, we calculated the mean of the maximum similarities obtained from each algorithm’s 100 runs.Biological interpretation for each module was provided based on the biological relevance of the phenotypes it mainly comprised, thereby generating the composite phenotypes.

Construction of multi-modal aging clocks

We constructed separate aging clock models for men and women across six organ systems, cognition, and the body. The organ systems clocks included the cardiovascular system, bone, kidney, metabolism, gut, and brain. The selection of aging biomarkers for each clock was determined by integrating prior knowledge^7–11,35 and considering the variability of individual phenotypes as well as composite phenotypes with age. Specifically, the biomarkers considered for each clock were as follows:

• Cardiovascular system: composite phenotypes M1 (ABI; baPWV; DBP; SBP; ankle DBP; ankle SBP; ankle pulse pressure; brachial DBP; brachial SBP; brachial pulse pressure; heart rate; QRS duration; QT duration; RV5; SV1), blood lipids (total cholesterol, triglycerides, LDL, HDL).

• Bone: composite phenotypes M15 (The BMD of several domains, including the lumbar spine, femoral neck, femoral neck, femoral shaft, trochanter, and ward).

• Gut: all the phenotypes of gut microbiome, except microbiome in M13 (Neither single phenotype nor composite phenotype was associated with age).

• Kidney: creatinine, uric acid, urea nitrogen, and cystatin C.

• Metabolism: all the phenotypes of the metabolome, except metabolites in M3 and M4 (Neither single phenotype nor composite phenotype was associated with age), selected phenotypes of blood biochemistry (glucose, total cholesterol, triglycerides, LDL, HDL).

• Brain: all the phenotypes of brain MRI.

• Cognition: all the phenotypes of the cognitive category (MMSE total score, MoCA total score, and six cognitive domains).

• Body: the phenotypes utilized in all the clocks mentioned above except gut microbiome (due to its limited sample size).

Each clock was constructed separately for men and women, with chronological age as the target variable and the selected biomarkers as predictors. The hyperparameters of support vector machine (SVM) considered were epsilon and cost, which were optimized using 10-fold cross-validation. The best model was selected based on its smallest error for the regression task and employed for predicting the chronological age. We obtained the age gap, which indicates variations in aging rates compared to peers, by regressing the predictive age against chronological age and calculating the residuals. The biological age was then derived by adding the age gap to the chronological age. The formulas could be shown as follows:

$$Y=X\beta +\mathrm{ϵ},$$

$$\mathrm{Age\; gap}=\mathrm{ϵ}=Y-X\beta ,$$

$$\mathrm{Biological\; Age}=X+\mathrm{ϵ}.$$

where $Y$ represented the predicted age, $X$ represented the chronological age, $\beta$ represented the parameters obtained from linear regression of chronological age on predicted age, and $ϵ$ represented the residuals (i.e., age gaps).

Exposure factors and health indicators

Factors can be categorized into the following groups:

• Sociodemographics: education year, number of family members, household income per capita.

• Lifestyle (variables were preprocessed to ensure that higher values corresponded to higher frequencies): current smoking (yes or no), current alcohol drinking (yes or no), current tea drinking (yes or no), sleep score.

• Diet pattern: overall plant-based diet (PDI), healthy plant-based diet (hPDI), unhealthy plant-based diet (uPDI).

• Menstrual cycle history: age at menarche, menses regular, menstrual cycle length, age at menopause.

• General functional health indicators: olfactory identification, gait score (TUG, Tinetti).

A harmonized sleep score was constructed based on self-reported sleep-related items common to all participants at baseline. As shown in Supplementary Table 3, the sleep parameters included snoring, insomnia, sleep duration, daytime alertness, and napping. Each item was binarized (1 = healthy response, 0 = unhealthy), and the total sleep score was obtained by summing across the five items, with higher scores reflecting better reported sleep. While this approach does not constitute a validated scale, it provides a pragmatic summary of multiple sleep-related dimensions³⁶ and aligns with methods used in previous large-scale cohort studies^37,38, although those studies included chronotype information, which was not captured in our questionnaire.

The three plant-based diet indexes (PDI, hPDI, and uPDI) were derived from a standardized food frequency questionnaire, which assessed the frequency and quantity of consumption over the past year for a range of major food items²³. To evaluate plant-based dietary patterns, 12 individual food groups were divided into three categories due to their varying health effects: healthy plant-based foods (whole grains, fresh fruits, fresh vegetables, and legumes), unhealthy plant-based foods (refined grains, sweets, and preserved plant-based foods), and animal-based foods (dairy products, eggs, seafood, meat, and processed animal-based foods). Using the grading system developed by Satiia et al.³⁹, each food group received either positive or reverse scores. In the scoring system, the top and lowest quintiles were reciprocally given scores of 5 and 1. Specific scoring criteria for various food groups can be found in the Supplementary Table 4. Ultimately, the scores from the 12 food groups were summed for each participant to create the three plant-based diet indexes. In our previous publication based on the same cohort, this method was used to construct the three indices, which were the primary exposures⁴⁰.

Menstrual cycle history was assessed through a structured questionnaire administered to all female participants. All data were self-reported retrospectively, as the participants were aged 55 years or older and thus overwhelmingly postmenopausal, with only four individuals reporting ongoing menstruation. Participants were asked to report their age at menarche (in years), whether they had experienced regular menstrual cycles during their peak reproductive years (i.e., between their twenties and early forties, prior to the peri-menopausal transition), and, if their cycles were regular, the typical number of days between two menstrual periods. Age at menopause was also recorded, defined as the age (in years) at which menstruation ceased permanently, following at least 12 consecutive months of amenorrhea.

Health status

We investigated the association between the biological age of various organs and three health status: metabolically unhealthy, osteoporosis, and cognitive impairment. The metabolic health status was categorized based on the cumulative count of metabolic risk factors (hypertension, hyperlipidemia, diabetes), with 0, 1, and 2–3 risk factors corresponding to metabolic healthy, moderately unhealthy, and highly unhealthy, respectively^41,42. Hypertension was diagnosed in individuals exhibiting a previous medical diagnosis of hypertension, a mean blood pressure of $\ge$140/90 mmHg, or the use of antihypertensive medications. The diagnosis of diabetes was established when there was a previous medical diagnosis, a fasting plasma glucose concentration of $\ge$7.0 mmol/L, or the use of anti-diabetic medications. For hyperlipidemia, a previous medical diagnosis was considered, along with criteria such as a serum CH concentration of $\ge$5.2 mmol/L, a TG concentration of $\ge$1.7 mmol/L, or the use of lipid-lowering medications. Participants with a BMD T-score of −2.5 or less in either the lumbar spine, total hip, or femoral neck were identified as having osteoporosis. Cognitive impairment assessment in the study utilized a stratified cut-off method based on education levels⁴³. For illiterate individuals, impairment was determined by MMSE$\le$17 and MoCA$\le$14. Among those with 0-6 years of education, impairment was identified with MMSE$\le$20 and MoCA$\le$20. For individuals with over 6 years of education, impairment was established by MMSE$\le$24 and MoCA$\le$25.

Statistics and Reproducibility

The age correlation of individual phenotypes and composite phenotypes were assessed. A two-sided Wilcoxon test was first conducted to evaluate sex differences in these phenotypes. Subsequently, Pearson partial correlation between chronological age and each individual phenotype were examined across all the participants, as well as separately among men and women, using the R package ppcor. In the overall cohort, sex was included as a covariate and total intracranial volume was further adjusted for brain MRI phenotypes; the same adjustment was also carried out in the sex-stratified analysis. Canonical correlation analysis (CCA) was employed to assess the relationship between these composite phenotypes and age in the entire population, as well as in men and women, using the function ‘cancor’ in the R package stats. All the multiple hypothesis testing results were adjusted using the false discovery rate (FDR) method.

The associations between multi-modal aging clocks and exposure factors were investigated using partial correlation analyses. Age and sex were included as covariates, and correlations between organ-specific age gaps and exposure variables were estimated, with results corrected for FDR. We employed mediation models to delve into the relationships between smoking, two aging clocks, and the olfactory identification ability, using the R package mediation with 1000 bootstrap resamples. Both two mediation models were adjusted for sex, age, current alcohol consumption, body mass index, hypertension, diabetes, and hyperlipidemia, and each model was further adjusted for organ-specific related factors that were found earlier. Specifically, the kidney model included additional adjustments for the number of family members and sleep score, and the brain model included additional adjustments for the number of family members. Covariates with missing values were imputed using the mean or median, as appropriate. All the multiple hypothesis testing results were adjusted using the FDR method.

The two-sided Wilcoxon test was employed to examine whether there were differences in age gaps of each organ between the normal and unhealthy groups, with FDR correction for multiple tests within each health status. The effect size was calculated as the Z statistic of the Wilcoxon test divided by the square root of the total sample size⁴⁴.

In the second follow-up, conducted at the median time of 6 years, participants with missing follow-up data were excluded from this longitudinal analysis. The occurrence of incident cardiovascular events was considered, including myocardial infarction, stroke, and intracerebral hemorrhage. We examined the association of the multi-modal biological ages as well as the chronological age with the risk of cardiovascular events using Cox proportional hazard models with the R package survival. The basic Cox models were conducted with adjustments for sex, while the extended models underwent further adjustments for additional covariates including current smoking status, current alcohol consumption, body mass index, hypertension, hyperlipidemia, diabetes and the use of antihypertensive, lipid-lowering, and glucose-lowering medications. Covariates with missing values were imputed using the mean or median, as appropriate.

Results

We harnessed the baseline data from the TIS, wherein the inclusion criteria comprised individuals aged between 55 and 65 years. This encompassed a cohort of 403 men and 501 women participants of Han Chinese descent (Supplementary Table 5). Drawing from a spectrum of twelve categories of phenotypes (Fig. 1a), we examined the age-related variations in individual phenotypes, extracted composite phenotypes through phenotype correlation networks, and further explored the relationship between composite phenotypes and age (Fig. 1b). Based on available phenotypes, we decided to establish a total of eight aging clocks, at three levels: organ systems, cognition, and whole body. The selection of phenotypes for each clock was based on prior knowledge and age-phenotype correlations we explored. The models were constructed using the SVM, resulting in the estimation of biological age. The age gap, representing the difference between biological and chronological age, was calculated, providing insight into the individual’s aging status in specific domains compared to same-sex peers of the same age. Subsequently, we probed the interplay and distinctions between these organ age gaps and exposure factors, as well as health status. Finally, biological ages were utilized as predictive tools for the second-round follow-up cardiovascular events risk (Fig. 1b).

Fig. 1. Overview of study design.

a Schematic of 12 categories of phenotypes derived from the Taizhou Imaging Study (TIS) at baseline. b Analysis pipeline. MRI, magnetic resonance imaging; SVM, support vector machine; BA, biological age.

Age-related single phenotype and composite phenotype

We investigated the variability of multi-scale phenotypes with age from individual and composite phenotypes’ perspectives. Given the significant differences in many phenotypes between sexes (Fig. 2a, Supplementary Data 1), we adjusted for sex when calculating the associations between phenotypes and age (Fig. 2b, Supplementary Data 3). Additionally, sex-stratified analyses were also conducted (Fig. 2c, Supplementary Data 4, 5). Across the entire cohort, 36 phenotypes exhibited a positive correlation with age, while 53 phenotypes displayed a negative correlation, after multiple testing adjustments with the FDR < 0.1. Individuals of advanced age were more likely to score lower on cognitive ability indicated by the MMSE, olfactory identification, as well as greater impairment in gait function indicated by the TUG test score. With increasing age, the most significant increment in brain MRI occurred in the volume of the left lateral ventricle, while the most substantial reduction was observed in the volume of the left thalamus. Elder individuals exhibited higher levels of lactic acid, LDL, and other blood analytes, while direct bilirubin levels decreased. The baPWV, ankle, and brachial pulse pressure showed a positive correlation with age, whereas ankle and brachial mean arterial pressure decreased. A decline in the abundance of several benefical gut microbiota species was observed with increasing age, such as Eubacterium hallii (metabolic versatility)⁴⁵, Anaerostipes hadrus (butyrate production)⁴⁶, and some from the Bacteroides genus (alleviate inflammation and intestinal injury)^47,48. Among these age-related phenotypes, 16 phenotype associations with age were consistently observed in sex-stratified analyses (Fig. 2b, c and Supplementary Data 3, 4, 5).

Fig. 2. Characterization of age-related single phenotypes and composite phenotypes.

a The proportion of phenotypes with significant differences between sexes in each category (determined by the two-sided Wilcoxon test, and adjusted with false discovery rate, FDR). b Phenotypes with significant positive or negative associations with age (Pearson partial correlation, two-sided t-test, P$<$ 0.1, FDR-corrected). In association analysis, we adjusted for sex, and brain MRIs were further adjusted for the estimated total intracranial volume. c Number of shared and specific phenotypes with positive or negative associations with age between all, men, and women participants, with shared phenotypes shown below. d Associations among all the phenotypes. For display purposes, only inter-categories significant associations were shown (Pearson partial correlation, age and sex adjusted, two-sided t-test, P$<$ 0.05, FDR-corrected). e The 15 composite phenotypes detected by the Louvain algorithm from the phenotypes association network. For display purposes, only edges with weights greater than 0.001, and those connecting phenotypes from different modules and categories were shown. c–e shared the same legend with distinct colors symbolizing different phenotype categories. f The modules’ stability and their correlation with age (canonical correlation analysis, two-sided Wilks’ Lambda test, corrected by FDR) within all, men, and women participants. P_adj, P value corrected by FDR method. MRI, magnetic resonance imaging; BMD, bone mineral density; baPWV, brachial-ankle pulse wave velocity; PP, pulse pressure; DBIL, direct bilirubin; MMSE, mini-mental state examination; MAP, mean arterial pressure; ECG, electrocardiography; ANTH, anthropometry. Statistical significance was denoted as follows: ns, not significant; +P$<$ 0.1, *P$<$ 0.05, **P$<$ 0.01, ***P$<$ 0.001. Exact P values and sample sizes for (a, b, f) are provided in Supplementary Data 1, 3, and 9, respectively.

Composite phenotypes were generated using unsupervised network-based module detection methods, which grouped statistically correlated variables into coherent modules^21,22. Fifteen modules were identified from the phenotype association network using the Louvain algorithm (Fig. 2d, 2e and Supplementary Data 6, 7, 8). Except for Module 7, the robustness of all other modules was greater than 0.6, indicating a relatively high level of stability under varying detection algorithms (Fig. 2f and Supplementary Data 9). Each module represented a cluster of phenotypes that were statistically correlated and may shared similar biological functions. These phenotype clusters were annotated based on the dominant biological characteristics of the included traits and then defined as composite phenotypes. The resulting composite phenotypes (Fig. 2e, Supplementary Data 7, 8) included cardiovascular system(M1), LDL subfractions (M2, M3, M4), HDL subfractions (M5), very low-density lipoprotein (VLDL) subfractions (M6, M7), biochemistry and anthropometry (M8), brain cortical average thickness (M9), brain region volume (M10), gray matter volume (M11), gut microbiome (M12, M13), cognition and gait (M14), and BMD (M15). The majority of the composite phenotypes demonstrated a high level of agreement with prior knowledge. For instance, M1 comprised electrocardiogram, vascular ultrasound, and blood pressure indicators. Highly diversified composite phenotypes, like M9, suggested interrelations among phenotypes such as brain cortical average thickness, olfaction, and metabolites. Notably, 11 modules revealed correlations with age after adjustments with FDR < 0.1 in the overall population (Fig. 2f and Supplementary Data 9). All single lipoproteins of metabolites didn’t show any age-related variance, whereas lipoprotein-based composite phenotypes (M2, M5, M6, M7) did. M1, M15, and the three brain composite phenotypes (M9, M10, M11) demonstrated a strong correlation with age. Although the gut microbiome composite phenotype mainly represented by M12, exhibited a high correlation with age, it did not reach statistical significance.

Multimodal aging clocks built from multi-phenotypes

Following prior knowledge of organ-specific biomarkers and the observed correlations with chronological age, phenotypes were retained for constructing organ-specific clocks only if either the individual traits or their composite phenotype showed significant age associations. Based on available phenotypes, a total of eight aging clocks were then developed at three levels: organ systems, cognition, and whole body, with separate constructions for men and women. Specifically, these models were developed for the cardiovascular system (n = 781, 339 men), bone (n = 855, 382 men), gut (n = 656, 290 men), kidney (n = 827, 361 men), metabolism (n = 827, 361 men), brain (n = 799, 346 men), cognition (n = 786, 350 men), and the body (n = 593, 251 men) (Supplementary Table 5). Among them, phenotypes from the M1 (cardiovascular) and M15 (BMD) composite phenotypes with strong biological coherence were used to construct corresponding organ-specific aging models. Phenotypes from M3 and M4 (LDL particle subfractions) and M13 (gut microbiota species) were excluded from the full set of metabolic and microbial variables used to construct the corresponding aging models, because neither individual phenotypes nor the composite phenotypes showed significant associations with chronological age. The body age model was built exploiting all phenotypes used in other domains, excluding the gut microbiome due to its limited sample size. The predictive accuracy for chronological age was highest in the predicted body (overall correlation coefficient, r = 0.866; mean absolute error, MAE = 1.605) age. Among various organs, the predicted age of the brain (r = 0.796; MAE = 1.853) showed the strongest correlation with chronological age, followed by metabolism, cardiovascular system, and gut (Fig. 3a and Supplementary Fig. 3).

Fig. 3. Characteristics of multi-modal aging clocks.

a The linear regression of predicted age derived from the prediction model against chronological age. The solid black dashed lines represent the best-fitting lines. b Density distribution of age gaps, with shaded areas indicating standard deviation (SD) ranges from the mean (zero). c, d UpSet plots showing the sample sizes of outliers across different organ domains. Outliers were identified as individuals with age gaps (difference between biological age and chronological age), exceeding ±2 SD from the mean. The bar chart on the right displays the number of outliers identified in each organ domain. The bottom matrix indicates the intersections of domains, and the bar chart above it represents the sample size corresponding to each intersection. c Outliers with accelerated aging. d Outliers with decelerated aging. e The network depicted the Pearson partial correlation (adjusted for age and sex, two-sided t-test, multiple testing adjusted by FDR) among organ ages (predicted ages after adjustment). Nodes represent organ ages, while edges represent partial correlations, with darker and thicker edges indicating stronger and more significant correlations, separately. f Heatmap of pairwise Pearson partial correlations (adjusted for age and sex, two-sided t-test, FDR-corrected) among organ age gaps. n: number of participants; r: Pearson correlation coefficient; MAE: mean absolute error. Cardiac, cardiovascular; Metab., metabolism; BA, biological age; Cor, correlation. The statistical significance was denoted as follows: ns, not significant; + P$<$ 0.1, * P$<$ 0.05, ** P$<$ 0.01, *** P$<$ 0.001. Exact P values and sample sizes for panels (e) and (f) are provided in Supplementary Table 6 and 7, respectively.

Next, the biological age gap, which reflects the state of aging speed among same-sex peers, was derived, along with biological age. The biological age gap ranged from 1.605 to 2.732 (Fig. 3b). A small proportion of participants had age gaps beyond ±2 standard deviations; these individuals were identified as outliers. Among the 455 participants with complete organ age gap data, the number of outliers per domain ranged from 7 to 15 (1.54%–3.30%) (Fig. 3c, d). Specifically, 15 individuals (3.30%) were accelerated aging outliers in the brain, and in the kidney, 9 (1.98%) showed accelerated and 14 (3.08%) showed decelerated aging outliers. Across domains, the brain and kidney were most frequently involved, with 3 individuals (0.66%) showing accelerated aging outliers in both.

To explore the interrelationships among multi-organ, cognition, and body aging, we examined partial correlations for their biological ages and biological age gaps separately (Fig. 3e, f; Supplementary Tables 6 and 7) and found that the significant connections (P $<$ 0.1, FDR-corrected) from both perspectives were quite similar. Whether considering biological age or age gap, the correlation coefficients amongst all aging clocks except the whole body remained below 0.25 (Fig. 3e, f; Supplementary Tables 6 and 7). Among all organs, the brain aging had the most connections with other organs, including the cardiovascular system, bone, kidney, and metabolism. Cognitive aging not only synchronized with brain aging (age cor =  0.152, P = 1.25 × 10⁻⁴; age gap cor = 0.084, P = 6.18 × 10⁻²; n = 705), but also exhibited a positive correlation with cardiovascular, gut and bone aging. Despite not incorporating the gut microbiome in the construction of the body model, body aging displayed a positive correlation with all other organs and cognition, including the gut. Among these, brain aging (age cor = 0.634, P = 2.31 × 10⁻⁶⁶; age gap cor = 0.609, P = 8.04 × 10⁻⁶⁰; n = 593) showing the strongest association with body aging (Fig. 3e, f; Supplementary Tables 6 and 7).

Diverse exposure factors related to distinct aging clocks

To explore how different aging clocks may reflect diverse aspects of aging, we evaluated their associations with a broad array of variables, including sociodemographics, lifestyle, and menstrual cycle history, and functional health indicators, adjusting for age and sex. After FDR correction, 10 factors were identified as significantly associated with the age gap across five aging clocks, while no associations were observed for the bone, metabolism, and gut clocks (P $<$ 0.1, Fig. 4a, Supplementary Data 10). Longer educational attainment and stronger olfactory identification ability correlated with decelerated cognitive aging. Meanwhile, the higher scores on the unhealthy plant-based diet index (uPDI) (r = 0.159, P = 1.30 × 10⁻⁴, n = 786) and TUG (r = 0.114, P = 3.37 × 10⁻², n = 620) were associated with accelerated cognitive aging. This linkage between cognitive aging and gait capability has been initially hinted at in the previously mentioned composite phenotype M14. Higher household per capita income tended to be linked with a more youthful body age than that of peers. Those who had a higher sleep score tended to exhibit a younger kidney (r = −0.088, P = 8.53 × 10⁻², n = 810) and brain (r = −0.103, P = 2.98 × 10⁻², n = 779) age gap compared to their peers. Smoking correlated with accelerated aging in the kidney (r = 0.201, P = 2.67 × 10⁻⁷, n = 824), brain (r = 0.122, P = 5.72 × 10⁻³, n = 796), and overall body (r = 0.102, P = 8.58 × 10⁻², n = 591). Apart from its association with younger cognitive age, a stronger olfactory identification ability was also aligned with a deceleration in aging across three other domains, specifically the kidney, brain, and overall body. The observed association between olfaction and brain aging aligned with the composite phenotype M9 mentioned earlier. Furthermore, we have identified connections between menstrual cycle history and aging clocks. A delayed onset of menarche was correlated with a younger cardiovascular, brain, and body age gap. Meanwhile, a longer menstrual cycle length was related to a more youthful brain and body age gap.

Fig. 4. Associations between aging clocks and exposure factors and health indicators.

a Heatmap displaying pairwise Pearson partial correlations (adjusted for age and sex, two-sided t-test, FDR-corrected) between the age gap of organs and two sets of variables: exposure factors and functional health indicators. b, c Estimated proportion of the association between current smoking status and olfactory identification mediated by the organ age gap. Mediation analyses were performed 1000 bootstrap resamples. Two-sided bootstrap tests were used to assess the significance of indirect effect (IE), direct effect (DE), and the proportion mediated. P values were not adjusted for multiple comparisons. Two models were adjusted for common covariates including sex, age, current alcohol consumption, body mass index, hypertension, diabetes, and hyperlipidemia, along with organ-specific exposure factors. b Kidney age gap, n = 629. The model was further adjusted for the number of (No.of) family members, and sleep score. c Brain age gap, n = 606. The model was further adjusted for the number of family members. Cardiac, cardiovascular; Cor, correlation; uPDI, unhealthy plant-based diet index; TUG, timed up and go test. The statistical significance was denoted as follows: +P$<$ 0.1, * P$<$ 0.05, ** P$<$ 0.01, *** P$<$ 0.001. Exact P values and sample sizes for panels (a–c) are provided in Supplementary Data 10, Supplementary Tables 8 and 9, respectively.

Given the positive relation of some age gaps with smoking and the negative link with olfactory identification, we further investigated whether there exist mediating relationships. The results revealed that the kidney age gap explained 6.94% (95%CI: 1.08% to 18.63%, n = 629) of the association between smoking and decreased olfactory identification scores, and the brain age gap explained 12.46% (95% CI: 4.37% to 24.44%, n = 606) (Fig. 4b, c and Supplementary Table 8, 9).

Connections between diverse health status and distinct aging clocks

Furthermore, we next aimed to explore whether the observed diversity in aging also corresponds with varied health status. Specifically, age gaps across aging clocks were compared between healthy and unhealthy groups for metabolic health, osteoporosis, and cognitive impairment. Among them, metabolic health status was classified by the count of metabolic risk factors (hypertension, hyperlipidemia, diabetes): 0 for healthy, 1 for moderately unhealthy, and 2–3 for highly unhealthy. Cardiovascular aging emerged as a direct reflection of metabolic unhealthiness, with the cardiovascular age gap significantly increasing alongside increasing metabolic risk (Fig. 5a, b, Supplementary Table 10, 11). Beyond cardiovascular aging, individuals in the metabolically unhealthy group demonstrated a slightly higher degree of kidney aging compared to those in the healthy group ($P=1.09\times {10}^{-1}$, effect size = 0.078). This difference was more pronounced in the subgroup with a high level of unhealthiness ($P=3.63\times {10}^{-2}$, effect size = 0.123). Moreover, when metabolic unhealthiness was high, the aging of the brain ($P=9.98\times {10}^{-2}$, effect size = 0.100) was notably greater compared to the healthy group, and the aging of the body ($P=1.42\times {10}^{-2}$, effect size = 0.083) showed a slightly greater compared to the moderately unhealthy group. Interestingly, metabolic age did not show a sensitive response to the current metabolic health status. In the context of osteoporosis, besides the bone age gap ($P=2.73\times {10}^{-2}$, effect size = 0.093) being significantly higher in the diseased group compared to the non-diseased group, the metabolic age gap ($P=2.73\times {10}^{-2}$, effect size = 0.100) contrasted by being relatively lower. Individuals with cognitive impairment exhibited higher levels of cognitive ($P=2.30\times {10}^{-5}$, effect size = 0.170) and body aging ($P=1.19\times {10}^{-4}$, effect size = 0.175) (Fig. 5a and Supplementary Table 10).

Fig. 5. The distribution of organ age gaps across three health status.

a The rows indicated the age gap of various organs, cognition, and body, while the columns represented the health status. The left and right violins corresponded to the normal and disease groups, respectively. Median and quartile information was displayed through the box plots, and the point denoted the mean of the age gap. Median comparisons of the age gap were conducted using the two-sided Wilcoxon test, and FDR-corrected. b The distribution difference of each organ age gap across metabolic health status, categorized by the number of metabolic risk factors (hypertension, hyperlipidemia, diabetes): 0 for metabolically healthy, 1 for moderately unhealthy, and 2–3 for highly unhealthy individuals. Boxplots show the median (line), interquartile range (box), and 1.5× IQR whiskers; The dot within each box represents the mean age gap, while outliers are shown as individual points beyond the whiskers. Median comparisons of the age gap were conducted using the two-sided Wilcoxon test with FDR correction. Cardiac, cardiovascular; Metab., metabolism. The statistical significance was denoted as follows: + P$<$ 0.1, * P$<$ 0.05, ** P$<$ 0.01, *** P$<$ 0.001. Exact P values and sample sizes for (a, b) are provided in Supplementary Table 10 and 11.

Prediction of cardiovascular events based on biological age

We next investigated whether organ age is a better predictor than chronological age for incident cardiovascular events six years later. Cox proportional hazards analyses demonstrated that organ ages were stronger predictors of incident cardiovascular events than chronological age in both sex-adjusted (Supplementary Fig. 4) and extended models (Fig. 6). In the extended model, controlling for traditional risk factors (sex, current smoking status, current alcohol consumption, body mass index, hypertension, hyperlipidemia, diabetes) and the use of antihypertensive, lipid-lowering, and glucose-lowering medications, biological age of the cardiovascular system (hazard ratio, HR = 1.14; 95% CI: 1.01–1.30), bone (HR = 1.15; 95% CI: 1.01–1.30), metabolism (HR = 1.17; 95% CI: 1.02–1.34), brain (HR = 1.14; 95% CI: 1.00–1.31) and overall body (HR = 1.18; 95% CI: 1.01–1.39) were significantly associated with cardiovascular risk. Notably, in each domain, BA yielded stronger effect estimates (i.e., higher HRs or narrower confidence intervals) than chronological age when modeled within the same sample set. In addition, models incorporating biological age for these specific organ systems showed higher Harrell’s C-index and lower Akaike information criterion values, indicating better overall predictive performance compared to models using chronological age (Fig. 6).

Fig. 6. Associations of organ ages or chronological age with 6 years follow-up cardiovascular events.

Cox proportional-hazards models were constructed with covariates including sex, current smoking status, current alcohol consumption, body mass index, hypertension, hyperlipidemia, diabetes, and use of antihypertensive, lipid-lowering, and glucose-lowering medications. Adjusted HR (95% CI) and P value (two-sided z-test) indicate the strength and significance of the association between BA or CA and cardiovascular events, after adjusting other covariates. P values were not adjusted for multiple comparisons. C-index and AIC reflect the overall predictive performance of the full model. For each comparison, adjacent models of the same color in the table adopted BA and CA as predictors, respectively, within the same set of samples. AIC, Akaike information criterion; BA, biological age; CA, chronological age; HR, hazard ratio; CI, confidence interval; C-index, Harrell’s C-index; Cardiac, cardiovascular; Metab., metabolism. Inc/Obs represented the number of incidents/the number of objects.

Discussion

In this study, we illuminated the heterogeneity of phenotypic aging and multi-organ aging, leveraging multi-scale phenotypes drawn from a late middle-aged population in rural China. Our exploration of phenotypic aging evolved from an examination of a single phenotype to composite phenotypes, helping us establish a hierarchy of biological clocks across various organ systems, cognition, and whole-body aspects. Notably, aging rates across various organs showed significant yet weak correlations, further emphasizing the interrelated yet diverse signature of organ aging. This diversity extended to varied modifiable factors linked with accelerated aging in each domain, as well as the related health status observed. In particular, accelerated aging of the brain and kidney played a mediating role in the decline of olfactory function at different levels. Finally, specific organ age proved to be a better indicator of cardiovascular events risk than chronological age.

Our findings in the age range of 55–65, captured phenotypic transitions from middle age to elderhood⁴⁹. Notably, we systematically revealed concurrent changes across multiple phenotypic domains, such as cognitive decline, olfactory disorder, gait impairment, decline in beneficial bacteria^50,51, and brain structural alterations. Unexpectedly, the minimal metabolic changes observed, particularly the absence of significant individual lipoprotein alterations, contrast with studies reporting pronounced changes in lipoprotein and hormone metabolism could occur mainly in the third and fifth decade in Chinese women⁵². This may reflect the narrow age range examined, where trends are not strictly linear. For instance, in Chinese men, total cholesterol rises before 40, plateaus at 41–60, then declines, while in women, it increases until 56–60 before decreasing⁵³. Thus, the relative stability of cholesterol levels in 55–65 may result from the counterbalance of these opposing trends. However, the age-related changes missed by individual lipoprotein markers were effectively captured by composite phenotypes, highlighting the importance of considering the joint behavior of multiple indicators in aging research.

Prior research on aging clocks has primarily focused on single determinants, such as DNA methylation⁵⁴ and omics approaches⁵⁵, while multi-organ studies have integrated diverse phenotypic markers. Nie et al. examined young and middle-aged Chinese adults (20–45 years), incorporating physiological, imaging, and omics data¹¹. Tian et al. applied a similar approach to UK adults (39–73 years)³⁵. Oh et al. extended this to U.S. cohorts (27–104 years) using plasma proteomics⁵. Their biomarker selection relied on organ-specific expression patterns or common knowledge attributing each marker to a specific system, such as brain MRI. Our study, focusing on the late middle-aged rural Chinese population, applied an unsupervised clustering approach to identify organ-related biomarkers from a wide array of phenotypic data. The rationale for using composite phenotypes lies in their capacity to reduce dimensionality while efficiently capturing integrative signals across related single phenotypes²². Notably, variables grouped into cardiovascular and BMD composite phenotypes were effectively used in downstream biomarker selection, demonstrating the potential of data-driven modules to reveal biologically coherent patterns that inform multi-organ aging models. From this perspective, composite phenotypes and aging clocks serve as complementary frameworks for elucidating the coordinated nature of aging across organ systems.

The unsupervised module detection did not always produce biologically coherent composite phenotypes, yet such heterogeneity proved valuable. Some modules, such as M9 and M14, captured cross-domain links, including those between brain structure and olfaction, and between cognition and gait. These seemingly inconsistent groupings echoed later in our findings: brain aging acceleration was associated with olfactory decline, and cognitive aging acceleration correlated with gait impairment. While unsupervised methods provide insight into latent biological structures, their utility in guiding practical biomarker selection remains a challenge. Notably, M1, which combined arterial ultrasound, blood pressure, and electrocardiography measures, consistently reflected cardiovascular function and proved valuable for organ age modeling. In contrast, most other modules lacked this level of integration and practical utility. This underscores the importance of combining unsupervised discovery with biological insight to construct robust and interpretable composite phenotypes.

We highlight the interconnected nature of aging across organ systems, with the brain playing a pivotal role. This was evident in the brain with other organs and the strong correlations observed between the brain age gap and that of other organs.These findings are align with research showing that accelerated biological aging in multiple body systems contributes to advanced brain aging, revealing directional relationships in aging pathways³⁵, as well as evidence that Alzheimer’s disease is associated with indicators of accelerated aging in 10 studied organs⁵. The brain, with its substantial demand for oxygen and nutrients⁵⁶, exhibits a heightened interdependence with the cardiovascular and metabolic systems. Kidney disease and brain structure are also highly interlinked. Prior research highlights the observation of cerebral vascular and neurodegenerative changes on MRI scans in patients with chronic kidney disease, which is frequently seen as a paradigm of expedited vascular brain aging^57–59. While the correlation of the bone with the brain might not be immediately apparent, existing studies suggest the possibility of the bone influencing the brain through various pathways^60–63. For example, osteocalcin, a bone-derived osteokine crossing the blood-brain barrier, is thought to accumulate in parts of the brain, potentially affecting brain development and cognitive functions^60,61. This intriguing interplay also aligns with our observations of a concurrent relationship between bone and cognitive aging. These explorations provide a more comprehensive understanding of multi-organ aging and emphasize the importance of overall health in maintaining the wellness of the brain and cognition.

Delving deeper, the patterns of aging vary considerably across organ systems. Notably, the brain exhibited the highest proportion of individuals identified as accelerated aging outliers, underscoring its potential sensitivity to pathology elsewhere in the body⁶⁴. In contrast, the kidneys showed a higher proportion of decelerated than accelerated aging outliers, indicating a unique resilience that may stem from their vast reserve capacity^65,66, allowing them to mitigate aging effects. Despite these speculations, understanding the complex aging process, shaped not only by organ resilience or susceptibility but also by intertwined genetic, environmental, and exposure factors, necessitates more longitudinal research to illuminate differences in aging across various physiological areas. Moreover, the relatively low levels of interconnections among organ age gaps are in line with previous research findings. For example, Tian et al. demonstrated that most of these correlations were below 0.1³⁵. Consistent with this, several other studies have also observed weak to moderate associations among different organs^5,11. Collectively, these results highlight the limited and variable inter-organ connectivity in the aging process.

Our exploration of organ-specific exposure factors further highlights the diverse influences that contribute to variations in aging across different organs. These could include modifiable lifestyle factors such as smoking cessation, improved income and education levels, better sleep quality, refined dietary patterns, and greater attention to menstrual health, etc. The further finding that smoking accelerates brain and kidney aging, which in turn contributes to impaired olfactory identification, provides a valuable addition to the well-established link between smoking and sensory decline^67,68. Highlighting these relationships may also improve public engagement with preventive behaviors, such as smoking cessation. While our study did not directly evaluate the modifiability of organ age gaps, the observed associations with lifestyle factors underscore the potential for these measures to serve as sensitive, dynamic markers in intervention studies. Given that the clinical outcome benefits of lifestyle (e.g., diet, sleep) or pharmacological interventions often require long follow-up, organ age models may offer an early, organ-level readout of intervention effects⁶⁹. This feature could facilitate more timely, targeted strategies in preventive and lifestyle medicine, supporting broader translational application.

We also identified multi-organ associations related to metabolic health, osteoporosis, and cognitive impairment. The greater the number of metabolic risk factors present, the more accelerated aging was observed in multiple organs. This may be attributed to the interplay of these risk factors^70,71, which potentially heightens their impacts, presents increased systemic inflammatory levels⁷², and highlights prominent signs of accelerated aging in organs such as the kidneys and brain^73,74. However, despite the critical role of metabolic health in these systems, the unresponsiveness of the metabolic age gap to metabolic status may be due to both the narrow age range, which fails to capture periods of significant metabolic fluctuations, and the limited scope of our metabolic measurements. Previous studies have reported higher metabolic age gap in conditions such as diabetes and chronic kidney disease among individuals aged 39–73³⁵. Additionally, while our dataset focused primarily on lipoproteins, it did not include other key metabolic components, such as fatty acids, their intermediates (e.g., acylcarnitines), and their downstream products (e.g., ketones), all of which play significant roles in established disease pathways^75,76. This limitation is particularly noteworthy given that fatty acids have been prominently featured in previous study as key indicators of metabolic age⁷⁷. Interestingly, we found that individuals with osteoporosis demonstrated slower metabolic aging, which is consistent with earlier research that illuminated the protective qualities of some lipoproteins toward bone density^78,79. Furthermore, individuals with cognitive impairments exhibited significantly accelerated body aging, while this is not reflected in brain aging. This might be because cognitive impairments can be caused by microscopic injuries to neurons or synapses⁸⁰, which were not captured by our current structural brain imaging technology⁸¹. Meanwhile, body aging, involving the cumulative effects of microscopic changes across various organ systems, might be more sensitively detected.

Our study further demonstrates that organ aging can predict cardiovascular events, offering potential directions for future research and clinical applications. Previous studies have found identified aging in the heart, kidney, artery, liver, and immune system is linked to an increased risk of heart failure^5,82. We also found that bone age is a significant predictor for cardiovascular diseases. This finding was not standalone; previous research has established correlations between bone health issues like bone loss and osteoporosis, and increased risks of cardiovascular diseases^83,84. Unlike traditional risk models, organ age offers a continuous and intuitive age-based scale that helps patients better understand their health status relative to peers of the same chronological age. This age-based framing may improve doctor–patient communication and promote adherence to preventive recommendations. When linked to disease risk, organ age can also be integrated into individualized assessments and targeted interventions. Differences in biological age across organs may help identify systems undergoing accelerated aging, guiding personalized lifestyle or pharmacological interventions, as well as informing routine health checkups or workplace wellness programs. Longitudinal monitoring of organ-specific age trajectories could provide quantitative feedback on intervention effectiveness and improve prediction of disease risk. However, the absence of a unified modeling framework, standardized thresholds linking organ age to risk, and consensus criteria for risk stratification currently limits the real-world applicability of these models. Additionally, their reliance on comprehensive, multi-dimensional data and the representativeness of the training cohort pose challenges to generalizability. Future efforts should prioritize standardized modeling protocols, validated risk thresholds, and external validation across diverse populations to enhance robustness and clinical adoption.

There are several limitations to this study. First, the cohort consisted of late midlife Han Chinese individuals from a single rural site, which may limit the generalizability of the findings to other ethnic, geographic, or populations across different age groups. Second, phenotypic aging was assumed to progress linearly within a relatively narrow age range, potentially missing non-linear patterns or critical transitions occurring earlier or later in life. Third, data-driven derived composite phenotypes aid in biomarker identification for high-dimensional data but do not always yield biologically coherent groupings. Their utility depends on integration with prior knowledge, and practical application remains a challenge. Fourth, the aging clocks were developed using cross-sectional data, which restricts the ability to capture individual-level aging trajectories and biomarker dynamics. Fifth, although the SVM model was applied for its simplicity and stability, it may not fully capture complex, high-dimensional interactions that could be revealed by more advanced algorithms such as multi-modal transformers¹⁴. Sixth, while a multi-organ approach was employed, certain domains were insufficiently covered; for example, the assessment of metabolic aging was constrained by limited biomarker availability, and brain aging was evaluated only through structural MRI. Expanding to include modalities such as comprehensive metabolic panels or functional neuroimaging⁸⁵ could enhance model depth. Seventh, sleep scores were not derived from standardized sleep scales, which may affect comparability and validity. Similarly, dietary scores were based on food frequency questionnaires, which are generally less accurate than 24- or 48-h recalls⁸⁶. Additionally, other variables obtained through self-reported questionnaires may be subject to recall bias, particularly menstrual history in postmenopausal women, which may affect their reliability. As such, their associations with aging metrics require further validation. Eighth, although organ age showed promise for disease risk prediction, clinically actionable thresholds could not be determined due to data limitations. The relatively limited number of incident disease events may also affect the precision of risk prediction estimates. Future studies with larger datasets and sufficient follow-up duration, combined with outcome-based model training, are needed to enable reliable threshold determination, enhance clinical translation and support practical application⁸⁷. Ninth, the study lacked external validation in independent cohorts, which may limit the robustness and generalizability of the findings. Tenth, as an observational study, causal inference cannot be established, and residual confounding cannot be fully excluded. Eleventh, while organ age offers a integrative measure, the biological mechanisms underlying organ-specific heterogeneity remain to be clarified. Finally, our study integrates diverse determinants of aging, though its complexity presents challenges. Future research with external cohorts and mechanistic investigations would strengthen its robustness and translational significance.

Conclusions

In conclusion, our study provides a comprehensive perspective on aging dynamics by uncovering synchronicity and heterogeneity at both phenotypic and organ levels, as demonstrated through phenotypic trajectories and organ-specific associations with health status, exposure factors, and cardiovascular events risk. We demonstrate the potential of unsupervised selection of multi-organ clock biomarkers and elucidate the interconnected yet divergent aging rates across different organs. Our findings highlight the role of organ aging including its mediation in functional decline, health status assessment, and disease prediction. Collectively, these insights advance our understanding of aging dynamics and contribute to the development of holistic geroprotective approaches that account for both combined and differential aspects of aging.

Author contributions

Conceptualization: Y.L., K.X., M.C., Y.J., and Y.W. Methodology: Y.L., X.X., Y.Z., X.Z., J.W., and J.L.L. Investigation: Y.L., R.L., N.G., Y.W., D.W., C.S., and T.Z. Visualization: Y.L., and J.W. Supervision: Y.L., K.X., X.C. Writing—original draft: Y.L., J.C.L., J.W., and X.X. Writing—review & editing: Y.L., X.X., K.X., X.C., and Z.L.

Peer review

Peer review information

Communications Medicine thanks Xueqiu Jian. Sajad Zalzala and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Data availability

The TIS data are available from the corresponding authors (xukelin@fudan.edu.cn; xingdongchen@fudan.edu.cn) upon reasonable request. Source data underlying Fig. 2 are provided in Supplementary Data 1, 3–9; Fig. 3 in Supplementary Tables 6 and 7; Fig. 4 in Supplementary Data 10, Supplementary Tables 8–9; Fig. 5 in Supplementary Tables 10–11. Data for Fig. 6 are contained within the figure itself.

Code availability

This study used open-source software R (version 4.4.1) and code. The R packages utilized are detailed in the Methods section. Code for conducting the core analyses is available on GitHub (https://github.com/LiyucanEDU/OrganAge).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01222-w.