Abstract
Biological age reflects actual ageing and overall health, but current
ageing clocks are often complex and difficult to interpret, which
limits their clinical application. This study introduces a Gompertz
law‐based biological age (GOLD BioAge) model designed to simplify the
assessment of ageing. We calculated GOLD BioAge using clinical
biomarkers and found significant associations between the difference
from chronological age (BioAgeDiff) and the risks of morbidity and
mortality in the NHANES and UK Biobank. Using proteomics and
metabolomics data, we developed GOLD ProtAge and MetAge, which
outperformed the clinical biomarker models in predicting mortality and
chronic disease risk in UK Biobank. Benchmark analyses demonstrated
that the models outperformed common ageing clocks in predicting
mortality across diverse age groups in both the NHANES and UK Biobank
cohorts. Additionally, a simplified version called Light BioAge is
created, which uses three biomarkers to assess ageing. The Light model
reliably captured the mortality risk across three validation cohorts
(CHARLS, RuLAS, and CLHLS). It significantly predicted the onset of
frailty, stratified frail individuals, and collectively identified
individuals at high risk of mortality. In summary, the GOLD BioAge
algorithm provides a valuable framework for the assessment of ageing in
public health and clinical practice.
Keywords: aging clocks, biological age, frailty index, metabolomics,
proteomics
__________________________________________________________________
The Gompertz law‐based biological age metric (GOLD BioAge) introduces
interpretable convenient and interpretable calculations to capture
mortality, disease, and frailty risks. Its applications to proteomics
and metabolomics (ProtAge and MetAge) demonstrate clinical potential to
enhance aging evaluation and prevent age‐related diseases. A simplified
version, Light BioAge, is developed and validated across independent
datasets.
graphic file with name ADVS-12-e01765-g002.jpg
1. Introduction
Human ageing manifests as progressive physiological changes and a
decline in physical and cognitive function that leads to an increased
risk of mortality.^[ [52]^1 ^] There is significant heterogeneity among
individuals during the ageing process,^[ [53]^2 ^] and chronological
age may not accurately reflect the actual pace of ageing. Furthermore,
because ageing is the primary risk factor for most chronic diseases,
targeting the ageing process itself may delay multiple age‐associated
diseases.^[ [54]^3 ^] Consequently, ageing assessments and treatments
have the potential to predict and prevent functional decline and
age‐related chronic diseases.^[ [55]^4 ^] Some routine clinical
biomarkers serve as biomarkers for ageing and predict the risks of
functional decline and mortality after adjusting for chronological
age.^[ [56]^5 ^] Integrating these biomarkers into composite panels may
offer a more comprehensive and powerful assessment of ageing than
single biomarkers alone.
Biological age measures an organism's biological functioning compared
with the expected level for a specific chronological age to reflect
overall health.^[ [57]^6 , [58]^7 ^] Levine's phenotypic age, which
integrates nine biomarkers with chronological age, can predict
mortality more accurately than chronological age alone.^[ [59]^8 ^]
Building on the concept of phenotypic age, Sheng et al. proposed PCAge
to estimate biological age through linear dimensionality reduction;
however, this method may be sensitive to outliers and thresholding
effects.^[ [60]^9 ^] Wei et al. presented ENABLAge, which integrates
machine‐learning models with explainable artificial intelligence to
ensure high prediction accuracy.^[ [61]^10 ^] In addition to clinical
ageing clocks, omics‐based ageing clocks hold significant promise as
they capture more precise dynamic molecular interactions and pathways
that are closely tied to the biological ageing process.^[ [62]^11 ^]
Specifically, epigenetic biomarkers have been utilized extensively in
DNA methylation ageing clocks, such as the Horvath Clock^[ [63]^12 ^]
and GrimAge Clock.^[ [64]^13 ^] Furthermore, multi‐tissue aging clocks
provide insights into the molecular changes that complex organisms
undergo with age to offer detailed information about ageing and
disease.^[ [65]^11 , [66]^14 , [67]^15 ^] Recently, proteomics and
metabolomics data from large cohorts have been used to accelerate the
development of plasma proteomic and metabolomic ageing clocks.^[
[68]^16 , [69]^17 , [70]^18 , [71]^19 , [72]^20 , [73]^21 ^] Proteomic
aging clocks show promising accuracy in predicting mortality and
multimorbidity,^[ [74]^14 , [75]^16 ^] Although these biological ageing
clocks demonstrate excellent performance in predicting disease and
mortality, however, their clinical translation remains limited because
of the gap between scientific research and their application in
clinical settings.^[ [76]^22 ^]
The complexity of current models of ageing clocks combined with
challenges related to interpretability, required features, and
generalizability may hinder their translation into clinical settings.
For example, despite the widespread adoption of DNA methylation
clocks,^[ [77]^12 ^] the collection of biological samples and reliance
on high‐throughput sequencing remain time intensive and costly. As
public demand for personalized health tracking grows, it becomes
critical to ensure the clinical relevance and practicality of ageing
clocks. Improvements to these tools must generate actionable clinical
insights while maintaining cost‐effectiveness, accessibility, and
robustness across diverse populations. Although omics‐driven methods
remain costly and technically demanding, aging clocks based on routine
clinical biomarkers offer a practical solution for health care
integration owing to their accessibility and widespread applicability.
Addressing these clinical challenges will require refining
computational methods and identifying robust biomarkers that balance
precision with practicality.
The Gompertz law is one of the most widely used mathematical models for
describing mortality. It effectively captures the exponential increase
in the mortality hazard across adult ages, in line with empirical
mortality data.^[ [78]^23 ^] The model's simplicity and flexibility
allow for its wide application. For example, Levine's phenotypic age
used the Gompertz model to estimate 10‐year mortality risk.^[ [79]^8 ^]
Additionally, Kuo et al. employed the model's cumulative mortality risk
to develop a proteomic ageing clock.^[ [80]^19 ^] The Gompertz model
provides a theoretical basis for biological ageing clocks in clinical
practice.
In this study, we develop a Gompertz law‐based biological age (GOLD
BioAge) algorithm that utilizes the hazard function of the Gompertz
distribution. This approach offers an easily calculated linear model
that combines chronological age and routine biomarkers to link the
deviation from chronological age to morbidity and mortality risks. We
apply the GOLD BioAge algorithm to metabolomics and proteomics data
from the UKB to investigate the algorithm's validity with omics‐based
data. Moreover, we compare its predictive performance for mortality
with common ageing clocks using data from the National Health and
Nutrition Examination Survey (NHANES) and the UK Biobank (UKB).
Finally, we refine and simplify GOLD BioAge as a light model and
validate it across three independent Chinese cohorts: the China Health
and Retirement Longitudinal Study (CHARLS), the Chinese Longitudinal
Healthy Longevity Survey (CLHLS), and the Rugao Longevity and Ageing
Study (RuLAS).
2. Results
2.1. Definition and Development of the GOLD BioAge Model
Biological age refers to the age that accurately reflects an
individual's risk of mortality. Higher mortality risk corresponds to
older biological age. Based on the Gompertz model, we linked
chronological age and biomarkers to mortality hazard with an
exponential distribution ([81]Supplementary Methods, Figure [82]1A).
Consequently, the Gompertz law‐based biological age (GOLD BioAge) was
estimated as the age that aligned with the joint mortality hazard
derived from both chronological age and biomarkers. Thus, the GOLD
Biological Age (GOLD BioAge) was derived as follows:
[MATH: GOLDBioAge=CA+∑βi
∗Biomaekeri+β0 :MATH]
(1)
with Biomarker[i] indicating the i‐th biomarker, β[ i ]as its
corresponding coefficient, and β[0] as the constant.
Figure 1.
Figure 1
[83]Open in a new tab
GOLD BioAge and its Association with Health‐Related Factors. Panel A
illustrates the exponential relationship between mortality hazard and
biological age (BA) and chronological age (CA). The “Diff” referred to
the difference between GOLD BioAge and CA, termed GOLD BioAgeDiff. The
scatter plot B) shows the strong correlation between GOLD BioAge
(estimated biological age) and CA. The estimated coefficients for CA
and biomarkers C), used to calculate GOLD BioAge, were displayed, with
the mean biomarker values of young adults serving as the reference.
Abbreviation: MCV: mean cell volume; RDW: red cell distribution width,
ALP: alkaline phosphatase, LYM%: lymphocyte percent, WBC: white blood
cell count, GGT: gamma glutamyl transferase (GGT). The distribution of
GOLD BioAgeDiff in NHANES D). The correlations of GOLD BioAgeDiff with
counts of age‐related chronic diseases E), self‐rated health F), and
unhealthy lifestyles (G).
The NHANES included 39,348 samples (49.5 ± 18.0 years old) with 26
biomarkers and chronological ages in the analysis. After feature
selection was implemented via LASSO‐Cox regression (Figure [84]S1,
Supporting Information), we developed a clinical ageing clock, GOLD
BioAge, based on chronological age and 9 biomarkers, which showed a
strong correlation with chronological age (R = 0.969, Figure [85]1B).
GOLD BioAge is the linear combination of chronological age, red blood
cell distribution width (RDW), albumin (ALB), creatinine, etc
(Figure [86]1C). It demonstrates how changes in specific biomarker
values, such as RDW and ALB, contribute to biological age to provide an
intuitive interpretation of biomarkers and ageing at the individual
level.
2.2. GOLD BioAgeDiff as a Novel Ageing Metric
We introduced GOLD biological age difference (BioAgeDiff) as the
difference between BioAge and chronological age to estimate the
magnitude of how individuals’ biological age deviates from their
chronological age (Figure [87]1A; Figure [88]S2, Supporting
Information). If the BioAgeDiff is greater/lower than 0, the person is
older/younger than the CA. The BioAgeDiff, as a linear combination of
biomarkers, establishes a clear relationship between changes in
biomarkers and shifts in biological age. This calculation of BioAgeDiff
facilitates understanding of how deviations in biomarkers from
reference values affect biological age. The BioAgeDiff can be
interpreted through the equation below:
[MATH: GOLDBioAgeDiffΔAge=∑βi<
mo>∗Biomarkeri−Biomarkerrefi+β0
′ :MATH]
(2)
where Biomarker [refi ]is the reference value of the i‐th biomarker.
For example, if an individual's blood glucose level increases by 1
mmol/L, BioAge increases by 0.58 years (Figure [89]1C).
Figure [90]1D shows the distribution of BioAgeDiff, which is close to a
normal distribution (mean: 0, standard deviation (SD): 5.57). When
major chronic diseases were considered, participants with comorbidities
had greater BioAgeDiff values than those without chronic diseases
(Figure [91]1E). Notably, individuals with four diseases were
approximately 5 years older in BioAge. With regard to health status, a
higher BioAgeDiff was found to be cross‐sectionally associated with
poorer self‐rated health (Figure [92]1F). Additionally, unhealthy
lifestyles, such as smoking and alcohol use, were associated with a
greater BioAgeDiff (Figure [93]1G). The results of BioAgeDiff were
validated in the UKB (Figure [94]S3, Supporting Information).
BioAgeDiff was associated with risks of mortality in the NHANES and UKB
(Table 1), with hazard ratios (HRs) of 1.155 (1.150–1.159) and 1.133
(1.131–1.135) per 1‐year increase, respectively. Survival curve
analysis (Figure [95]2 ) of 20‐year follow‐up data revealed that
participants in the highest 25% of the BioAgeDiff groups had a steeper
decline in survival probability than those in the lowest 25% of the
groups, especially among middle‐aged and older age groups. For example,
among individuals aged 65–74 years, approximately 74.9% of those in the
high‐risk group died after approximately 16 years, whereas only
approximately 30.2% of those in the low‐risk group died. BioAgeDiff can
be considered a measure through linear dimension reduction or
projection. Thus, we also compared the performance of BioAgeDiff with
common metrics, including the Mahalanobis distance statistic^[ [96]^24
, [97]^25 ^](MDS) and principal component analysis^[ [98]^26 ^] (PCA).
Among middle‐aged (45–64 years) and older (65–85 years) age groups,
BioAgeDiff outperformed other linear metrics in identifying individuals
at high risk of mortality (Figure [99]2).
Figure 2.
Figure 2
[100]Open in a new tab
The associations of BioAgeDiff and risks of mortality. Survival plots
for individuals categorized by BioAgeDiff, PCA age, and MDS in the
NHANES cohort are presented. The high and low risk groups represent the
top and bottom 25% of the age‐stratified population (ages 45–54, 55–64,
65–74, and 75–85 years). PCA: principal component analysis; MDS:
Mahalanobis distance statistics.
2.3. Application of GOLD BioAge to Metabolomics and Proteomics
We applied our algorithm to create the MetAge and ProtAge models on the
basis of blood NMR metabolomics and proteomics data in the UKB,
respectively. Like the clinical‐based BioAge, the omics‐based ageing
clocks showed strong correlations with chronological age and
age‐related factors (Figure [101]3A; Figure [102]S4, Supporting
Information). ProtAge exhibited a significant ability to capture
mortality risk that surpassed MetAge, clinical BioAge and chronological
age (Figure [103]3B). For all‐cause mortality, ProtAge achieved a
C‐index of 0.790, whereas MetAge and BioAge reached 0.747 and 0.738,
respectively. These results were consistent across different age groups
and cause‐specific mortality rates (Table [104]S11, Supporting
Information). Notably, among young adults (<45 years), ProtAge had a
C‐index of 0.793 in survival analysis, highlighting its effectiveness
in predicting the risk of premature mortality (Figure [105]3C). For
cause‐specific mortality, ProtAge had a C‐index of 0.754 for cancer
mortality and 0.850 for heart disease mortality, the highest among the
three ageing clocks. Compared with those with MetAgeDiff and
BioAgeDiff, individuals in the top 25% according to ProtAgeDiff
presented the highest cumulative mortality incidence rates throughout
the follow‐up period (Figure [106]S5, Supporting Information). Pathway
enrichment analysis of ProtAge proteins (Figure [107]S5, Supporting
Information) identified regulation of MAPK cascade (HGF, GDF15, EGFR,
etc.), and regulation of intracellular signal transduction (REN, AGER,
KIT, ect.), highlighting their potential roles in regulating cellular
senescence traits.^[ [108]^27 ^]
Figure 3.
Figure 3
[109]Open in a new tab
The associations of GOLD ProtAge, MetAge, and BioAge with mortality in
UK Biobank. A) Correlations between the three aging clocks and
chronological age. B) C‐index values from survival analysis and the AUC
for 10‐year mortality prediction, comparing the three aging clocks and
chronological age, with results for all‐cause (age‐stratified) and
cause‐specific mortality. C) Survival curves for individuals classified
by ProtAgeDiff, MetAgeDiff, and BioAgeDiff in the general population
(top panel) and young adults (bottom panel, <45 years old). High and
low risk groups are defined as the top and bottom 25% of the
population. ProtAgeDiff, MetAgeDiff, and BioAgeDiff represent the
differences between ProtAge, MetAge, and BioAge and chronological age,
respectively. The C‐index for ProtAgeDiff and its subpanels and
proteins are shown. ProtAgeDiff consisted of CardioDiff, InfamDiff,
NeuroDiff, and OncoDiff, which were linear combinations of
cardiometabolic, inflammatory, neurological, and oncological proteins.
E) Density plots and G) a correlation heatmap (filled with Pearson
correlation coefficients) of these subpanels are presented. F) Survival
plots based on ProtAgeDiff subpanels and H) the risk score, which was
the count of high‐risk factors derived from ProtAgeDiff subpanels.
We then decomposed ProtAgeDiff into contributions from cardiometabolic
(CardioDiff), inflammatory (InflamDiff), neurological (NeuroDiff), and
oncological (OncoDiff) proteins (Figure [110]3D,E), which may reveal
various aspects of ageing mechanisms. CardioDiff and NeuroDiff emerged
as the top two contributors to ProtAgeDiff and demonstrated the highest
C‐index in survival analysis (Figure [111]3F; Figures [112]S6,S7,
Supporting Information). Among these proteins (Table 2), GDF15,
NTproBNP, and EGFR have been identified as ageing biomarkers, and NEFL
is frequently highlighted among neurological proteins. Given the
relative independence of the four ProtAgeDiff categories
(Figure [113]3G), we used the counts within the high‐risk group (top
25% of Cardio/Neuro/Inflamm/Onco Diff) to produce a risk score ranging
from 0 to 4. This risk score effectively identified individuals at high
risk of mortality (Figure [114]3H). For example, approximately 60% of
people who scored 4 died within approximately 16 years due to all‐cause
mortality. In summary, ProtAge and its ProtAgeDiff serve as omics‐based
ageing clocks for predicting mortality risk, and the ProtAgeDiff
calculation allows us to analyse the ageing process across four
distinct biological categories.
2.4. GOLD BioAge and Incident Chronic Diseases
To investigate the potential of GOLD BioAge to predict the incidence of
common chronic diseases, we included cancer, myocardial infarction
(MI), heart failure, stroke, chronic obstructive pulmonary disease
(COPD), and dementia in the association analysis. The Cox proportional
hazards model showed that a 1‐year increase in biological age was
associated with an increased risk of disease (Figure [115]4A). For
example, in the case of cancer, a 1‐year increase in ProtAge, MetAge
and BioAge was associated with increases of 2.7%, 1.8%, and 1.9% in
hazard ratios (HRs), respectively. This trend was consistent across
other diseases, such as myocardial infarction and stroke. Moreover, the
ProtAge model demonstrated slightly higher HRs and C‐index values for
most specific diseases than the BioAge model did. For dementia, for
example, the HR of ProtAge reached 1.078 (1.069–1.087) per 1‐year
increase, whereas BioAge had an HR of 1.049 (1.045–1.054). Similarly,
the MetAge model exhibited robust performance across diseases such as
MI and stroke, with HRs of 1.082 (1.077–1.086) and 1.066 (1.061–1.071),
respectively. These results highlight the value of ProtAge and MetAge
in predicting incident chronic diseases in large cohorts.
Figure 4.
Figure 4
[116]Open in a new tab
Associations between GOLD ProtAge, MetAge, and BioAge and the incidence
of age‐related chronic diseases. The forest plots A) illustrates the
hazard ratios and C‐index for ProtAge, MetAge, and BioAge across
chronic diseases in the UKB. These associations were adjusted for age
and sex. MI: myocardial ischemia; COPD: chronic obstructive pulmonary
disease; CI: confidence interval. Survival plots are displayed based on
the differences between ProtAge, MetAge, and BioAge relative to
chronological age, referred to as ProtAgeDiff, MetAgeDiff, and
BioAgeDiff. The high and low risk groups correspond to the top and
bottom 25% of the population, respectively.
Cumulative disease incidence trajectories are presented on the basis of
the pace of ageing as measured by ProtAgeDiff, MetAgeDiff, and
BioAgeDiff (Figure [117]4B). The differences between the highest and
lowest ProtAgeDiff groups were most pronounced among the three metrics,
indicating that ProtAge was particularly effective in predicting the
onset of chronic diseases. Over a follow‐up period of 16 years, the
cumulative incidences for cancer, myocardial infarction, heart failure,
stroke, chronic obstructive pulmonary disease (COPD), and dementia in
the high‐ProtAgeDiff group were 27.0%, 11.6%, 5.4%, 7.9%, 16.1%, and
8.6%, respectively. Overall, these findings underscore the potential of
ProtAge and MetAge to predict age‐related chronic diseases.
2.5. Comparison with Other Ageing Clocks
To investigate the validity of our models, we compared the mortality
prediction performance of the GOLD BioAge model with Levine's
phenotypic age, KDM biological age (KDM‐BA), and chronological age in
the NHANES (8,106 participants, aged 47.0 ± 16.3 years) and UKB
(265,541 participants, aged 56.5 ± 8.0 years). These ageing clock
models were constructed using clinical biomarkers with chronological
age included as a reference.
Figure [118]5 shows the C‐index of survival analysis and AUC values for
10‐year mortality prediction of these ageing clocks. The BioAge model
performed significantly better overall than other biological and
chronological age models across the NHANES and UKB datasets, both in
the overall samples and within specific age groups. For example, the
BioAge model achieved a C‐index of 0.829 in the NHANES, outperforming
Levine's phenotypic age (0.827), KDM (0.807), and chronological age
(0.802). For cause‐specific mortality, the GOLD BioAge model had the
highest C‐index and AUC values among these ageing clocks. Taking
mortality related to heart disease as an example, the C‐index of the
BioAge was 0.881 in the NHANES and 0.780 in the UKB. We used the NHANES
III dataset as the validation dataset (Figure [119]S8, Supporting
Information), and the GOLD BioAge showed competitive performance
compared with these common ageing clocks. These results confirmed the
validity of our biological age algorithm and its ability to capture
mortality risk.
Figure 5.
Figure 5
[120]Open in a new tab
Comparison of GOLD BioAge and other common aging clocks in predicting
mortality in NHANES and UKB. The C‐index in survival analysis A) and
AUC value of 10‐year mortality prediction B) of these aging clocks are
shown. Both all‐cause (age‐stratified) and cause‐specific mortality are
considered. The highest value is marked with bold. The BioAge, Light,
Levine and KDM referred to the GOLD BioAge, its light version, Levine's
phenotypic age, KDM algorithm derived age, respectively.
2.6. Light BioAge for Practical Simplicity
For simplicity in clinical practice, we refined and simplified GOLD
BioAge as a light version called Light BioAge (Figure [121]S9,
Supporting Information). The Light BioAge model incorporates
chronological age, serum creatinine, glucose, and log‐transformed
C‐reactive protein (Log CRP). The calculation formula is as follows:
[MATH: Light
BioAge=Age+8.3313∗Creatinin
e+0.8270∗Glucose
+5.7305∗LogCRP−13.5298 :MATH]
(3)
In the NHANES, 38,001 samples (49.6 ± 18.3 years old) with these three
biomarkers were included. Light BioAge was strongly correlated with
chronological age (R = 0.989, Figure [122]6A), which accounted for
93.73% of the variance in the GOLD BioAge. Its difference from
chronological age (Light BioAgeDiff) was positively correlated with age
(Figure [123]6B), which followed a nearly normal distribution
(Figure [124]6C). It was also significantly associated with
comorbidities, self‐rated health, unhealthy lifestyles, and the risk of
mortality (Figure [125]6D–G, Table 1).
Figure 6.
Figure 6
[126]Open in a new tab
The Light BioAge and its associations with age‐related factors and
outcomes. The correlation of Light BioAge with age A). And the
difference of Light BioAge with age also correlated with age B) and its
distribution C). The correlations of Light BioAgeDiff with counts of
age‐related chronic diseases D), self‐rated health E), and unhealthy
lifestyles F). The survival plot G) based on Light BioAgeDiff levels,
with the top and bottom 25% of the population representing the high and
low risk groups. The forest plots H) show the hazard ratios and C‐index
of Light BioAge in relation to chronic diseases, adjusted for age and
sex. MI: myocardial ischemia; COPD: chronic obstructive pulmonary
disease; HR: hazard ratio; CI: confidence interval.
Compared with the GOLD BioAge model, the Light BioAge model, which
utilizes the fewest indicators, demonstrated competitive predictive
accuracy (Figure [127]5). In the NHANES dataset, while the full BioAge
model achieved a higher C‐index of 0.829 for all‐cause mortality, the
light model demonstrated competitive performance with a C‐index of
0.810. Furthermore, the C‐index of Light BioAge was very close to
previous prominent measures, such as Levine's phenotypic age and KDM.
For example, for mortality from cerebrovascular disease, the C‐index of
Light BioAge reached 0.911, which is comparable to the phenotypic age
(0.903) and KDM (0.914) reported in the NHANES. Notably, to enhance its
clinical applicability, we evaluated its performance in predicting
incident chronic disease. For example, the Light BioAge model
demonstrated HRs of 1.122 (1.116–1.128), 1.096 (1.090–1.102), and 1.062
(1.055–1.069) for COPD, MI, and stroke, respectively (Figure [128]6H).
These results indicate that Light BioAge provides a robust and
practical alternative while remaining competitive with other ageing
metrics.
2.7. Light BioAge Predicted Mortality in Validation Cohorts
We further validated Light BioAge in three independent datasets, the
CHARLS (17,163 participants, aged 58.4 ± 10.05 years), RuLAS (1,785
participants, aged 77.0 ± 4.2 years), and CLHLS (2,499 participants,
aged 85.5 ± 12.0 years) datasets. In the three cohorts (Table 1), 1752,
186, and 813 deaths occurred during the median follow‐up periods of
9.0, 4.0, and 4.1 years, respectively.
Light BioAge was strongly correlated with chronological age across the
three cohorts (Figure [129]7A). In the full samples, Light BioAge
achieved AUC values of 0.794 in the CHARLS, 0.809 in the CLHLS, and
0.753 in the RuLAS (Figure [130]7B). These values were greater than
those for chronological age, which were 0.778, 0.790, and 0.645,
respectively. Notably, Light BioAge outperformed chronological age for
individuals aged 60–79 years, with an AUC that exceeded 0.790 in both
the CLHLS and RuLAS. It also maintained an approximately robust AUC of
0.8 for those aged 80 and older, significantly outperforming
chronological age. Participants with high BioAgeDiff (top 25%)
experienced a more pronounced decline in survival probability than did
those with low BioAgeDiff (bottom 25%) across the CHARLS, RLAS, and
CLHLS (Figure [131]7C). By the end of the follow‐up periods in each
cohort, the survival probabilities of individuals in the high‐risk
groups were approximately 78.5.%, 84.5%, and 55.8%, respectively.
Figure 7.
Figure 7
[132]Open in a new tab
Validations of the Light BioAge in three independent cohorts. The
correlations A) of Light BioAge with age in CHALS, RuLAS and CLHLS. The
ROC curves B) of Light BioAge (solid lines) and age (dotted lines) for
predicting mortality across all samples, and within age‐stratified
groups (<80, ≥80 years old). Survival plots C) depict mortality
trajectories of individuals categorized based on Light BioAgeDiff
levels, with the top and bottom 25% represented as high and low risk
groups in CHARLS, RuLAS, and CLHLS.
Considering human ageing as a longitudinal process, we examined the
dynamic changes in Light BioAgeDiff between wave 1 and wave 3 of the
CHARLS (Figure [133]8A). Light BioAge in the two waves was strongly
correlated (R = 0.915, Figure [134]8B), whereas the Light BioAgeDiff
showed a moderate correlation (R = 0.475). In accordance with the Light
BioAgeDiff, participants were classified into slow (Diff < 0), normal
(0 ≤ Diff < 5) and fast (Diff > 5) ageing groups, which were
subsequently classified into seven categories on the basis of their
ageing status across both waves (Figure [135]8C). The stable
slow‐ageing groups across the two waves were used as the reference.
Compared with the reference group, the stable fast‐aging and
accelerated aging groups (slow/normal to fast) presented the highest
mortality risk (Figure [136]8D,E). In addition, the decelerated ageing
group (fast to slow/normal) was associated with a reduced risk of
mortality.
Figure 8.
Figure 8
[137]Open in a new tab
The Light BioAge, its dynamics and mortality in CHARLS. Illustration A)
detailing the study designs across five waves in CHARLS. Correlation B)
between Light BioAge values in wave 1 (2011) and wave 3 (2015). Scatter
plot C) displays Light BioAgeDiff in wave 1 (2011) and wave 3 (2015),
with dotted lines indicating Light BioAgeDiff values of 0 and 5.
Individuals were divided into 7 groups based on the changes in Light
BioAgeDiff, with survival plots D) and forest plots E) provided. Model
1 represented the crude model, while Model 2 adjusted for age and sex.
2.8. Light BioAgeDiff, Frailty and Mortality Risks
Next, we explored the associations of Light BioAgeDiff with frailty as
assessed by the frailty index, which included age‐related chronic
diseases, self‐rated health, basic and instrumental activities of daily
living and mobility capacity. In the CHARLS 2011 and 2015, frailty
status was associated with BioAgeDiff, in which frail individuals were
1.14 and 1.20 years older, respectively, than their robust counterparts
(Figure [138]9A). During longitudinal follow‐up (2011‐2015, 2015–2018),
the baseline BioAgeDiff was associated with incident frailty (odds
ratio [95% CI]: 1.03 [1.01‐1.04]; 1.04 [1.01‐1.07], Figure [139]9B).
The participants in the fourth quartile of BioAgeDiff presented the
highest risk. Using BioAgeDiff as a measure of biological ageing, we
examined the mediating role of functional decline, measured by the
frailty index, in the associations of BioAgeDiff with mortality risk
(Figure [140]9C). The frailty index mediated approximately 26.4% (p <
0.001) of the relationship with an increase accounting for 6.48% of the
variance.
Figure 9.
Figure 9
[141]Open in a new tab
The Light BioAgeDiff, frailty and mortality in CHARLS. The boxplots A)
of Light BioAgeDiff across robust, prefrail, and frail individuals in
CHARLS waves 1 (2011) and 3 (2015), with statistical significance
determined using Wilcoxon tests. Forest plots B) illustrates the
associations between Light BioAgeDiff and incidence of frailty. The odd
ratios were calculated through continuous and category Light BioAgeDiff
(Q1‐4: quartiles), adjusted by age and sex. The mediation models C) of
Light BioAge (wave 1, 2011), frailty index (wave 3, 2015) and mortality
(wave 3–5, 2015–2020). The change in frailty index was calculated based
on assessments from waves 1 and 3. SRH: self‐rated health; ADL:
activities of daily living; ADE: average direct effect; ACME: average
causal mediated effect. The survival plots of individuals according to
frailty status D), and both frailty status and levels of Light
BioAgeDiff E).
Light BioAgeDiff exhibited a performance similar to that of frailty in
predicting mortality, with C‐index values of 0.634 for BioAgeDiff and
0.633 for frailty in the CHARLS study. We used BioAgeDiff and frailty
as measures of biological and functional ageing and examined their
combined effectiveness in identifying individuals at high risk of
mortality. The mortality rate of frail individuals was approximately
29.2% during the 9‐year follow‐up according to the CHARLS
(Figure [142]9D). In contrast, frail individuals with the highest
BioAgeDiff had a mortality rate of approximately 40.1% during this
period (Figure [143]9E). These findings highlight the potential role of
Light BioAgeDiff in preventing incident frailty and its joint
contribution with frailty in identifying individuals at elevated risk
for mortality.
3. Discussion
In this study, we introduced a robust algorithm to estimate biological
age as a linear combination of chronological age and various
biomarkers. GOLD biological age and its difference from chronological
age provide insights into the relationship between individual biomarker
values and the pace of ageing. Notably, the implementation of our
algorithm in proteomics and metabolomics demonstrated the significant
potential of omics biomarkers in identifying the risks of mortality and
age‐related chronic diseases. Furthermore, benchmark analysis
demonstrated that our models outperformed traditional ageing clocks in
predicting the risks of both all‐cause and cause‐specific mortality
across different age groups. We also developed a simplified version
called Light BioAge, which provides a practical and efficient
alternative with simplified calculations. Light BioAge exhibited
predictive capabilities in assessing mortality risk across three
validation cohorts of elderly participants and was associated with the
onset of frailty. In summary, our algorithm was validated as a general
framework for constructing ageing clocks. Importantly, both GOLD BioAge
and its light version can serve as convenient tools for ageing
assessment in clinical practice.
The robustness of the GOLD BioAge algorithm and the ageing clocks was
validated through multiple aspects. The evaluation of GOLD BioAge
focused primarily on the correlation between BioAgeDiff and
chronological age as well as its ability to predict all‐cause and
cause‐specific mortality, the incidence of multiple age‐related chronic
diseases, the onset of frailty, and validations across diverse
populations. Benchmark analyses of mortality prediction demonstrated
the superiority and sensitivity of the GOLD BioAge model. In summary,
the GOLD BioAge serves as a general measure of biological ageing and
offers simple and practical calculations for ageing assessment and
public health.
The pace of individual ageing undergoes dynamic changes throughout life
and is influenced by modifiable lifestyles, environmental factors,
psychological influences, and health conditions. Identifying
individuals at high risk of premature ageing can increase primary
prevention efforts and reduce the health care and socioeconomic burdens
linked to age‐related diseases. In this study, GOLD BioAge estimated
individuals’ biological ageing status and captured the risks of
morbidity and mortality. To further promote the application of
biological age in public health and clinical settings, we introduced
Light BioAge, a simple and practical ageing clock that utilizes only
three accessible biomarkers in addition to chronological age. Light
BioAge demonstrated applicability across various independent cohorts
(NHANES, UKB, CHARLS, CLHLS, and RuLAS) with diverse study designs,
participant characteristics, and morbidity profiles. This model
incorporates serum creatinine, blood glucose, and C‐reactive protein
levels with chronological age to reflect kidney function and metabolic
and inflammatory status. These biomarkers are commonly used in medical
examinations and are easily accessible at low cost. Therefore, Light
BioAge offers a convenient tool for ongoing monitoring of ageing
trajectories to prevent functional decline and age‐related diseases.
Compared with Levine's phenotypic age, we estimated biological age by
fitting Gompertz mortality hazard function to empirical mortality data.
Levine's phenotypic age has been widely used in ageing‐related studies.
Notably, phenotypic age outperformed earlier biological age methods in
predicting all‐cause mortality and various diseases.^[ [144]^28 ^] GOLD
BioAge exhibited a strong correlation with Levine's phenotypic age in
the NHANES and UKB datasets, confirming the robustness and reliability
of our algorithm. Notably, Levine's phenotypic age relies on the
Gompertz cumulative distribution function to estimate the 10‐year
mortality risk. Its calculation involves a double logarithmic
transformation, which complicates its clinical interpretation. In
comparison, the calculation of GOLD BioAge is simplified by employing
the hazard function to identify the instantaneous mortality risk.
In addition to ageing clocks based on clinical biomarkers, our study
introduced ProtAge and MetAge, which incorporated omics data into the
GOLD biological age framework. These ageing clocks generally outperform
clinical marker‐based clocks in predicting mortality, which may be due
to the higher sensitivity and coverage of omics data in capturing
ageing‐related information.^[ [145]^29 ^] For ProtAge, proteins are
categorized into four groups on the basis of their physiological
function, with each group contributing to a distinct age estimate.
Protein expression and posttranslational modifications, particularly
those linked to inflammation, oxidative stress, and cell cycle
regulation, provide stable, long‐term biomarkers for clinical outcomes.
Additionally, because protein alterations often precede the onset of
chronic diseases, proteomics enhances early disease detection, making
ProtAge a valuable tool for predicting mortality and early‐stage health
risks.[146] ^30 , [147]^31 , [148]^32 ^] Metabolomics reflects rapid,
short‐term fluctuations in the body's biochemical processes, which
offers insights into the effect of recent changes in diet, physical
activity, and stress on ageing. By integrating both proteomic and
metabolomic data into the ageing clock, we created a more comprehensive
tool for estimating biological age. The GOLD omics ageing clocks
provide the potential for personalized health interventions to mitigate
ageing‐related risks.
Organ‐specific ageing clocks,^[ ^] such as brain age measured through
magnetic resonance imaging (MRI) data^[ [149]^34 , [150]^35 ^] have
been developed to evaluate biological ageing in individual organs.
Proteomic data further contribute to this understanding by including
organ‐specific proteins, which predict chronic diseases linked to those
organs.^[ [151]^14 , [152]^36 ^] However, research reveals that ageing
trajectories differ markedly across organs, with distinct patterns that
deviate from one another and from the body's overall rate of ageing.^[
[153]^33 , [154]^37 ^] These observations emphasize the need to enhance
the GOLD BioAge framework to address organ‐level variations in ageing
and better capture these biological disparities in the future.
This study has several limitations. First, although omics‐based ageing
clocks demonstrated superior performance compared with those that used
clinical biomarkers in the UKB dataset, further validation in other
elderly cohorts is essential to confirm these findings. Additionally,
biomarkers for ageing clocks were selected via LASSO penalized
regression to increase accuracy; however, different feature selection
methods could yield alternative sets of biomarkers, indicating the
potential for further optimization of biomarker panels in clinical
applications. While Light BioAge and GOLD BioAge demonstrated potential
as practical clinical tools, ProtAge and MetAge, as advanced
research‐oriented models, lacked immediate clinical applicability due
to their technical complexity and reliance on specialized data.
Furthermore, we validated Light BioAge in three Chinese cohorts, but it
is uncertain whether the full GOLD BioAge model would more accurately
capture the risks associated with geriatric syndromes and mortality.
4. Experimental Section
Study Populations
This study used the data of the NHANES 1999–2018, UKB, CHARLS, CLHLS,
and RuLAS, with baseline participant characteristics summarized in
Tables [155]S1,S2 (Supporting Information). The US NHANES was a
nationally representative cross‐sectional survey of civilians living in
the US that was approved by the National Center for Health Statistics
(NCHS) Ethics Review Board.^[ [156]^38 ^] The UK Biobank was a
large‐scale prospective cohort that collects data from over 500,000
participants across 22 centres in England, Scotland, and Wales. The UKB
received ethics approval from the North West Multicentre Research
Ethics Committee.^[ [157]^39 ^] The CHARLS was an ongoing
population‐based longitudinal cohort study of middle‐aged and older
Chinese adults. It was approved by the Ethics Review Board of Peking
University in accordance with the Declaration of Helsinki and other
relevant guidelines and regulations.^[ [158]^40 ^] The CLHLS was a
nationwide longitudinal study of the elderly Chinese population. The
project was approved by the Biomedical Ethics Committee of Peking
University, China (IRB00001052‐13074).^[ [159]^41 ^] The Rugao
Longevity and Ageing Study (RuLAS) was a population‐based prospective
study in Rugao, China, that consists of a longevity cohort and an
ageing cohort.^[ [160]^42 ^] The RuLAS was approved by the Human Ethics
Committee of Fudan University School of Life Sciences. All participants
provided written informed consent. This study followed the
Strengthening the Reporting of Observational Studies in Epidemiology
(STROBE) reporting guidelines for cohort studies.^[ [161]^43 ^]
Clinical Biomarker Selection for Constructing GOLD BioAge
Data was utilized from the NHANES 1999–2018 for variable selection and
refined the biomarker panel to construct the GOLD biological age. A
total of 26 common biomarkers from cell blood count (CBC) tests and
biochemical assays were included (Table [162]S3, Supporting
Information). LASSO‐Cox regression models were employed to select
biomarkers for predicting all‐cause mortality, with fivefold
cross‐validation to determine the optimal lambda value (lambda =
0.0166). Among the initial 26 biomarkers, 9 biomarkers were retained
(Figure [163]S1, Supporting Information). This set of biomarkers formed
the basis for the biological age model (GOLD BioAge). To make the panel
more practical for use, feature selection was conducted on a set of 10
blood (biochemical and haematological) biomarkers that were
consistently collected across the aforementioned five cohorts (Table
[164]S4, Supporting Information). This simplified panel, including
chronological age, serum creatinine, glucose, and C‐reactive protein
(CRP), formed the Light GOLD Biological Age Model (Light BioAge).
Metabolomics and Proteomics Biomarker Selection
GOLD BioAge models were applied to metabolomic and proteomic
biomarkers, employing data from the UK Biobank (UKB, 2006–2010). For
the GOLD Proteomic Age Model (GOLD ProtAge), 2,923 proteins from 53,014
participants were analyzed. For quality control, proteins were excluded
with more than 10% missing data and removed participants with more than
50% missing proteins, resulting in 1,459 protein biomarkers for feature
selection. A LASSO‐Cox regression model was then used with fivefold
cross‐validation optimized for all‐cause mortality prediction. To
optimize both model performance and simplicity, for a lambda value
(model penalty) of exp (‐4) was opted, which selected fewer features
while maintaining a relatively high C‐index (Figure [165]S1, Supporting
Information). Finally, 22 protein biomarkers associated with
chronological age were selected. The final dataset included 39,772
participants for downstream analysis, after excluding participants with
missing data of these 22 proteins (Table [166]S5, Supporting
Information). For the metabolomics data in UKB, a total of 248,202 UKB
participants were enrolled, each with measurements of 251 circulating
metabolomic markers. Similar to protein feature selection, LASSO Cox
regression was conducted for all‐cause mortality via fivefold
cross‐validation. Consequently, 26 metabolomic biomarkers were selected
on the basis of the lambda value of exp (‐6), which corresponded to an
increase of one standard deviation over the lambda value with maximum
C‐index. Detailed descriptions of all selected variables were available
in the supporting Information. The robustness of Lasso‐Cox
regression‐derived biomarkers was systematically assessed through
stability testing against data perturbations and subsampling strategies
(Supporting Information). The sensitivity analyses demonstrated
consistent selection of biomarkers, confirming their robustness to
perturbations and subsampling variability.
GOLD BioAge Model Training
Two Gompertz regression models were conducted for biological age model
training. The first Gompertz regression model included only
chronological age as a predictor of time‐to‐mortality data. The second
Gompertz regression model incorporated both chronological age and
selected biomarkers as predictors. GOLD biological age (GOLD BioAge)
was defined as age accounting for the actual mortality hazard by
considering both chronological age (CA) and additional biomarkers
(Figure [167]S2, Supporting Information). The models were specified as
follows:
Model 1: chronological age only:
[MATH: h1t=rate1∗expβ1∗CA+s
hape1∗t :MATH]
(4)
Model 2: chronological age and selected biomarkers:
[MATH: h2t=rate2∗expβ2∗CA+∑β2<
mi>i∗Biom
arkeri+shape2∗t :MATH]
(5)
In Model 2, biomarkers represented the selected biomarkers included in
the model, where Biomarker[i] was the i‐th biomarker and β[2i
]represented the coefficient of each biomarker in model 2. GOLD BioAge
integrates chronological age with relevant biomarkers to better
estimate mortality hazard and ageing status. Let h [1](Bioage, t = 0)
≈ h [2](CA,Biomarkers, t = 0). In the real dataset, the empirical
values of h [1] and h [2] were slightly different (Figure [168]S2,
Supporting Information) but strongly correlated (Pearson r = 0.969). To
reduce estimation bias for biological age in the whole population, a
variable (γ) was added to correct bias and let h [1] = γ*h [2] and
[MATH:
γ=h1h2 :MATH]
.
Thus, GOLD BioAge was derived as follows:
[MATH: GOLDBioAge=1β1<
/msub>β2∗CA+∑β2<
mi mathvariant="normal">i∗Biom
arkeri+logγ∗rate
2rate1 :MATH]
(6)
In the equation, log denoted the natural logarithm. For further
simplify the formula of GOLD BioAge, the coefficient parameter of CA
(β[2 ]) was set equal to the parameter (β[1]) in Model 2, and estimated
the parameters for the rate, shape, and coefficients of biomarkers in
Model 2. When β[2 ] = β[1 ] , the formula was further simplified as
follows:
[MATH:
GOLDBioAge=CA+∑β
2iβ1
∗Biom
arkeri+1β1
∗logγ∗rate
2rate1 :MATH]
(7)
The Gompertz distribution parameters (rate, shape, and coefficients)
were estimated by maximum likelihood using the “flexsurv” R package.
The constant item,
[MATH:
1β1<
mrow>∗logγ∗
rate2rate1
:MATH]
, was estimated by the mean value of
[MATH:
1β1<
mrow>∗logh<
mn>1h2∗
rate2rate1
:MATH]
, where h [1] and h [2] were the empirical values of hazard risk for
the two Gompertz models. The detailed coefficients of GOLD BioAge,
Light BioAge, ProtAge and MetAge were shown in Tables [169]S6–S9
(Supporting Information). This methodological validation included
direct comparison of Gompertz versus Cox regression frameworks for
biological age derivation ([170]Supplementary Methods). The biomarker
coefficients exhibited high concordance between approaches, driving
perfect alignment (R = 1.00) between GOLD BioAge and Cox BioAge
constructs (Figure [171]S10, Supporting Information). The algorithm of
GOLD BioAge was implemented as an R package
([172]http://github.com/Jerryhaom/GOLDBioAge). The GOLD BioAge tools
were also accessible through an online web calculator
([173]https://jerryhaom.github.io/GOLDBioAgeWeb.io/).
Benchmark of Biological Age Models
To ensure the robustness of the models, their performance in the NHANES
and UKB was assessed compared with other established phenotypic ageing
clocks and dimensional reduction methods. Levine's phenotypic age, KDM
age and the Mahalanobis distance statistic were calculated using the
“BioAge” R package.^[ [174]^44 ^] PCA age was calculated through
principal component analysis with the first five components regressed
to age. For benchmarking analyses, NHANES and UK Biobank (UKB) samples
were restricted to participants with complete biomarker profiles
required for the aging clock models. Predictive performance was
assessed through survival model discrimination (C‐index) and binary
classification accuracy (AUC) for 10‐year mortality, with stratified
analyses conducted across: the full cohort, age‐stratified subgroups:
young adults (<45 years), middle‐aged adults (45–64 years), and older
adults (>65 years), and sex‐specific subgroups (Tables [175]S10,S11,
Supporting Information). Additionally, the ability of these ageing
clocks were examined to predict cause‐specific mortality. Since each
ageing clock served as a single variable, logistic regression models
were used to estimate 10‐year survival predictions solely on the basis
of these ageing clocks and age. The C‐index of survival analysis (crude
model) was used to evaluate the discriminatory power of these ageing
clocks in predicting all‐cause and cause‐specific mortality. This
comprehensive benchmarking analysis provided a thorough evaluation of
the models' performance and facilitated comparisons with other
established ageing clocks.
Assessment of Mortality and Onset of Chronic Diseases
In the NHANES, death information was based on linked data from records
taken from the National Death Index (NDI) through December 31, 2019,
provided through the Centers for Disease Control and Prevention. Data
on mortality status and length of follow‐up (in person‐months) were
available for nearly all participants. In the UKB, death information
was obtained from death certificates held within the National Health
Service (NHS) Information Centre (England and Wales) and the NHS
Central Register (Scotland) to November 30, 2022. Participants’ time
was calculated to death from baseline to the date of death, date of
loss to follow‐up, or date of last record of follow‐up, whichever came
first. The International Statistical Classification of Diseases, 10^th
edition was used, to define causes of death. Cause‐specific mortality
included mortality from malignant neoplasms, heart disease,
cerebrovascular disease, respiratory disease, Alzheimer's disease,
diabetes, and others. In addition, data on the dates of incident
chronic disease in the UKB, including cancer, myocardial infarction,
heart failure, stroke, chronic obstructive pulmonary disease (COPD),
and dementia, were collected. The associations of GOLD BioAge, Light
BioAge, ProtAge, and MetAge with mortality and diseases were shown in
Tables [176]S12–S15 (Supporting Information).
Assessment of Health‐Related Factors and Outcomes
The unhealthy lifestyle score was based on six modifiable lifestyle
factors, namely, smoking, alcohol consumption, physical activity, diet,
body mass index (BMI), and sedentary behavior, as defined by the World
Health Organization. The score was categorized into five groups (0, 1,
2, 3, 4 and more unhealthy factors). Multimorbidities were defined as
the number of lifetime disease diagnoses. In the NHANES, diabetes, high
blood pressure, congestive heart failure, coronary heart disease, heart
attack, stroke, cancer or malignancy, and chronic bronchitis were
included. In the UKB, cancer, myocardial infarction, heart failure,
stroke, chronic obstructive pulmonary disease (COPD), and dementia was
included. Disease count was classified into five categories: no disease
and 1, 2, 3, and 4 or more diseases. Self‐rated health was recorded on
four levels: excellent or very good, good, fair and poor. The
distributions of GOLD BioAge, Light BioAge, ProtAge, and MetAge by
unhealthy lifestyle, comorbidity, and self‐rated health were shown in
Table [177]S16 (Supporting Information).
Validation in Independent Elderly Cohorts
The Light BioAge model was validated in three additional elderly
cohorts: the CHARLS, RuLAS and CLHLS datasets (Table [178]S2,
Supporting Information). Five waves of CHARLS data (2011–2020) were
utilized, and blood‐based bioassay data (CHARLS 2011, 2015) were used
to construct the Light BioAge. Health and function questionnaires were
collected for frailty assessment^[ [179]^45 ^] (Table [180]S17,
Supporting Information). For the RuLAS, wave 2 (2016) was used as the
baseline, and blood biomarker data were obtained. The CLHLS 2014–2018
data were used to validate Light BioAge. Using the R package ‘gbm’, we
applied gradient boosting models to survival time‐to‐event data to
train Light BioAge and age as mortality predictors. ROC curves were
calculated to evaluate the prediction performance of Light BioAge and
chronological age across all samples as well as subpopulations
stratified by age (60‐79 years; ≥ 80 years). Additionally, survival
curves were fit for the low‐risk and high‐risk groups on the basis of
the Light BioAgeDiff model.
Statistical Analysis
Survival analysis was conducted in different age groups. Within the
same group, participants were classified into quartiles (Table
[181]S18, Supporting Information) based on their BioAge Difference
(BioAgeDiff), with the top 25% representing individuals at the highest
risk of death. Kaplan‐Meier survival curves were then plotted to
compare the predicted survival probabilities between the highest and
lowest quartiles of the BioAge. Harrell's Concordance Index (C‐index)
was applied to quantify the discrimination accuracy of survival models,
while the Area Under the Curve (AUC) served as a metric to evaluate
predictive performance for 10‐year mortality status within a binary
classification framework. Cox proportional hazard models were conducted
to assess the associations between different biological aging clocks,
mortality and the onset of chronic diseases. The cox regression models
were adjusted for sex and chronological age. The pearson correlation
coefficient was used to quantify the correlations. The mediation
analysis was conducted through R package ‘mediaton’, through the
bootstrap approach. All statistical analyses were performed using R
version 4.3.3.
Conflict of Interest
The authors declare no conflict of interest.
Author Contributions
M.H., H.Z., J.W. contributed equally to this work. M.H., L.Y., H.Z.
performed concept and design. M.H., Z.H., S.J. performed acquisition,
analysis, or interpretation of data. M.H., J.W., H.Z. draft the
manuscript. All authors performed critical revision of the manuscript
for important intellectual content. M.H., H.Z., J.W. performed
statistical analysis. X.L., S.W., M.W., Y.H., J.W., J.C., Z.B., L.J.
performed administrative, technical, or material support. M.H., Y.L.,
S.J., Z.H., X.W. performed supervision
Supporting information
Supporting Information
[182]ADVS-12-e01765-s003.pdf^ (7.8MB, pdf)
Supplementary Table
[183]ADVS-12-e01765-s001.xlsx^ (53KB, xlsx)
Supporting Information
[184]ADVS-12-e01765-s002.docx^ (62.3KB, docx)
Acknowledgements