Introduction

Aging is widely regarded as a risk factor for the development of numerous chronic diseases1,2,3,4. Since aging is a multifactorial process influenced by both environmental factors (e.g., diet, exercise, and the physical environment) and genetic determinants5,6,7,8, quantifying the aging rate poses considerable challenges. This complexity raises an intriguing question: how old are we? Although determining age on the basis of birthdate seems straightforward, individuals with the same chronological age (CA) can exhibit strikingly different physical appearances. This phenomenon highlights the limitations of CA as a reliable indicator of the aging process. Numerous studies have consistently demonstrated that individuals have a biological age (BA), a measure of an individual’s physiological and functional state, reflecting the cumulative effects of aging processes, disease influence, and genetic factors, which may differ from their CA, that often diverges from their CA9,10,11. With advances in molecular biology and multiomics technologies, long-term scientists have observed molecular changes associated with the aging process. This has prompted the question of whether these aging-associated changes could be used to predict BA. As research into aging mechanisms continues to expand12,13,14,15, scientists have recognized that aging can be quantified using biomarkers. This insight has inspired the development of aging clocks designed to accurately measure BA.

The development of machine learning has brought renewed optimism to the field of aging clocks. These models can construct aging clocks by identifying patterns in molecular features across large datasets, including CpG methylation levels and protein concentrations11,16,17. In 2013, Steve Horvath developed the first multitissue predictor of age using data from 8000 samples derived from 82 DNA methylation array datasets16. In the same year, Carlos Lopez-Otin et al. systematically summarized the nine hallmarks of aging for the first time18. A decade later, Carlos López-Otín and colleagues updated the hallmarks of aging, expanding the list to twelve, with the addition of the hallmark of loss of proteostasis7. As direct biological effectors, plasma proteins undergo continuous dynamic changes throughout the aging process, as increasingly supported by research19,20. Tanaka et al. developed the first protein-based aging clock21, and in 2024, Austin Argentieri and colleagues demonstrated that proteomic aging clocks are also associated with mortality and lifespan.

The first proteomics aging clock, developed by Toshiko Tanaka, was used to analyze proteomic data from 240 healthy females and males, identifying 197 proteins that were positively correlated with age and 20 proteins that were negatively correlated with age21. Building on this foundation, Sathyan et al. extended this work by profiling 4265 plasma proteins in older adults, identifying 754 proteins linked to CA and health outcomes22. Around the same time, the other team conducted a large-scale study revealed nonlinear proteomic remodeling across the lifespan and identifying 491 proteins highly predictive of aging19. Later, Wang et al. expanded the field by quantifying 4712 plasma proteins in nearly 12,000 participants from Atherosclerosis Risk in Communities, linking proteomic age acceleration to chronic diseases and mortality23. Most recently, Kuo et al. developed optimized proteomic clocks that predict BA, multimorbidity, and mortality risk with high precision24. In 2024, M. Austin Argentieri used plasma proteins to measure BA and predict the risk for aging-associated diseases, multimorbidity and mortality25. In our study, we aim to elucidate sex-specific differences in BA and mortality, and to systematically profile and characterize the proteins that contribute to these disparities.

The UK Biobank (UKB) is a large-scale biomedical database established in 2006 that comprises data from 53,013 participants across the UK at the time of recruitment. The UKB Pharma Proteomics Project (UKB-PPP) is a collaborative extension of the UKB that specifically focuses on proteomic profiling, detected by Olink platform. The Olink platform uses Proximity Extension Assay (PEA) technology as detailed in previous studies26,27. Considering the extensive proteomic data and the related population information available in the UKB-PPP dataset, we constructed a proteomics-based aging clock using the UKB-PPP dataset.

As described by Jing-Dong J. Han, the age-independent difference between the individual’s predicted BA which is predicted by molecular, cellular, or physiological biomarkers, and CA values is called ΔAge28. A positive ΔAge denotes “accelerating agers” (BA > CA), while a negative ΔAge denotes “decelerating agers” (BA < CA). In this study, we observe significant sex differences through distinct ΔAge-based aging trajectories between females and males. In females, the aging trajectory shows more nonlinear changes than in males. To identify proteins that modulate biological aging rates, we impose three selection criteria: significantly up- or down-regulated in “accelerating agers” versus “decelerating agers”, having a non-zero, positive/negative coefficient in the ProteAge aging-rate model, and exhibiting a continuous aging-related change (identified via Mfuzz clustering). Integrating these criteria, we screen two sets of aging-associated protein signatures accelerating aging rate proteins (AARPs) and decelerating aging rate proteins (DARPs). Compared with the male cohort, we identify more AARPs in the female cohort, while the number of DARPs shows no significant sexual differences. Finally, we evaluate how these AARPs and DAPRs predict mortality, and identify a subset of mortality-aging-associated proteins. We perform sex-specific mortality analysis for all populations, and the number of mortality-aging-associated proteins (n = 172, such as GDF15 and IL1R2) identified in the male cohort is significantly higher than those identified in the female cohort (n = 1, ZNRD2). Most AARPs act as risk factors in the mortality analysis; however, ITGAV, although classified as an AARP in aging-related analysis, is considered a protective factor in the mortality analysis, suggesting its complex role in the aging process.

Methods

UKB cohort

We collected clinical and plasma proteomics data from the UKB database29, which included data from 53,013 participants between 2006 and 2010. The UKB has collected comprehensive phenotypic and genotypic information from its participants, including data derived from questionnaires, physical measurements, sample assays, accelerometry, multimodal imaging, genome-wide genotyping, and longitudinal follow-up on a wide array of health-associated outcomes. Ethics approval for the UKB was obtained from the North West Multi-centre Research Ethics Committee (MREC, https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/about-us/ethics), and all participants provided written informed consent30. This research was conducted utilizing the UKB resource under approved application number 99709.

The UKB-PPP represents a collaboration between the UKB and 13 biopharmaceutical companies, yielding plasma proteomic profiles over 54,000 participants, individuals with available age information and over 80% proteomics data completeness were included for subsequent analysis30. Plasma was isolated from blood samples collected in EDTA (9 mL) vacutainers, which were subsequently stored at −80 °C before Olink analysis. The Olink platform quantified 2923 unique proteins by utilizing PEA combined with next-generation sequencing (Olink Explore 3072). The Olink platform employs PEA technology, utilizing dual antibody probes tagged with unique complementary DNA oligonucleotides. These proximity probes specifically bind to their target protein epitopes, inducing spatial convergence that facilitates oligonucleotide hybridization. This molecular interaction creates a amplifiable DNA template that proportionally reflects the original protein concentration. Subsequent quantification is achieved through next-generation sequencing (NGS)-based detection, a methodological approach whose technical specifications have been thoroughly characterized in prior publications26,27. The Olink measurements, including normalization, were based on the NPX values recommended by the manufacturer. The quality control details are provided at the following link: biobank.ndph.ox.ac.uk/ukb/ukb/docs/PPP_Phase_1_QC_dataset_companion_doc.pdf.

Definition of chronological age

CA was calculated using data from the UK Biobank. Specifically, we utilized the date of attending the assessment center (Field ID 53), along with the participants’ year (Field ID 34) and month (Field ID 52) of birth. CA was calculated using the following formula: the time difference between the assessment date and the participant’s birth month and year, divided by 12 to convert to years.

KDM and PhenoAge

We calculated BA with two different algorithms, the KDM31 and PhenoAge32. Both algorithms are implemented in the BioAge R package (version 0.1.0), which was developed by Kwon D and Belsky DW, and the details of the KDM (Klemera-Doubal method) and PhenoAge were described previously33. The KDM algorithm derives BA from regressions of clinical biomarkers on CA, whereas PhenoAge calculates BA on the basis of a mortality prediction score that incorporates clinical biomarkers and CA.

Owing to the flexibility of the BioAge package, we can customize BA algorithms with any aging-related biomarkers. The KDM algorithms include nine biomarkers34: forced expiratory volume (L), systolic blood pressure (mmHg), blood urea nitrogen (mg/dL), glycohemoglobin (%), total cholesterol (mg/dL), creatinine (μmol/L), albumin (g/L), alkaline phosphatase (U/L), and C-reactive protein (mg/dL). The PhenoAge algorithm also includes nine biomarkers: creatinine (µmol/L), serum glucose (mmol/L), red cell distribution width (%), albumin (g/L), alkaline phosphatase (U/L), mean cell volume (fL), lymphocyte percentage (%), C-reactive protein (mg/dL), and white blood cell count (1000 cells/µL). We obtained biomarker and demographic data from the NHANES III, a cross-sectional, population-representative survey conducted by the US Centers for Disease Control and Prevention between 1988 and 1994 (n ≈ 33,994 adults aged 20–80). We employed the NHANES functions of the BioAge package to train the KDM and PhenoAge algorithms in the NHANES III database and subsequently utilized the _calc functions to project these algorithms onto the UKB data.

ProteAge

A proteomics prediction model, ProteAge, was developed using the glmnet R package (version 4.1-8). We randomly selected 70% of the UKB-PPP data as the training set for the model. To ensure optimal robustness, we tuned two parameters: α (the penalty weight) and λ (the complexity of the penalty). After leave-one-out cross-validation using the cv.glmnet function (nfold = 20), we selected lambda.min (known as the λ of the minimum mean cross-validated error) as the final λ value. For the parameter α, we tested 20 values ranging from 0 to 1 in increments of 0.05 and selected the α value with the minimum MAE. We subsequently projected the trained algorithm to the remaining 30% of the UKB-PPP data for downstream analysis.

Time-series analysis of protein expression

For time-series expression pattern analysis, the male cohort was further divided by single-year CA from 40 to 70 years. For each age group, the mean expression level of each protein was calculated for downstream analysis. The optimal number of clusters was estimated using the minimum centroid distance method (Dmin function) across a range of cluster numbers (2–10), with three repetitions per setting. Subsequently, fuzzy c-means clustering was performed using the mfuzz algorithm with a fixed number of clusters (k = 5) and a predefined random seed to ensure reproducibility. The clustering results were exported for further interpretation.

Mortality prediction

Death status of individual (date of death) was conducted through integration with national death registries in the UK, incorporating administrative records from NHS Digital (England and Wales) and NHS Central Register (Scotland). The data harmonization process and validation protocols are documented in central UK Biobank death linkage documentation (https://biobank.ndph.ox.ac.uk/ukb/ukb/docs/DeathLinkage.pdf). We constructed three separate multivariable Cox proportional hazards regression models to assess the associations between individual biological aging measures and mortality risk. Each model included CA and one standardized BA residual (per standard deviation) as predictors. Cox regression models followed the standard guideline of at least 10 events per variable (EPV) to ensure stability and avoid overfitting35. Therefore, only individuals aged 40 and above were included due to insufficient sample size below this threshold. We further stratified participants into 10-year age groups to ensure statistical power. Specifically, the models were specified as follows:

  1. 1.

    Model 1: CA and ProteAge ΔAge /SD

  2. 2.

    Model 2: CA and KDM ΔAge /SD

  3. 3.

    Model 3: CA and PhenoAge ΔAge /SD

The models were fit using the coxph() function from the R package survival. Hazard ratios (HRs) and 95% confidence intervals (CIs) were obtained by exponentiating the regression coefficients. Model outputs were summarized using the tbl_regression() function (with exponentiated results) and the summary() function.

The hazard function for each model was formulated as:

$$h\left(t\right)={h}_{0}\left(t\right)\exp ({{{\rm{\beta }}}}_{1}\times {{\rm{CA}}}+{{{\rm{\beta }}}}_{2}\times {{\rm{BA}}}\, \Delta {{\rm{Age}}}/{{\rm{SD}}})$$

where, h(t) is the hazard at time t, h0(t) is the baseline hazard, CA represents chronological age, and BA ΔAge/SD represents the standardized BA residual derived from ProteoAge, KDM, or PhenoAge. HRs quantify the relative risk of mortality associated with a one SD increase in BA Δage, holding CA constand. An HR > 1 indicates increased mortality risk per unit increase in ΔAge, while an HR < 1 indicates a protective association. Living individuals were right-censored in this study, and the survival duration was calculated as days between the date of attending assessment center and the date of death ascertainment or last follow-up.

Differential expression analysis of proteomics data

For the proteomics data obtained from the Olink platform, we conducted differential expression analysis using the limma R package (version 3.58.1) The detailed methodologies have been described previously (https://github.com/UK-Biobank/UKB-RAP-Notebooks-Access). Proteins were considered differentially expressed if they met both of the following criteria: a false discovery rate (FDR) < 0.05 and an absolute log₂ fold change (|log₂FC|) greater than log₂(1). This combined threshold ensures both statistical significance and biological relevance.

Mortality-aging-associated proteins

To identify mortality-aging-associated proteins, we employed three strategies to classify proteins into AARPs and DARPs. First, we selected proteins that exhibited significant expression differences between accelerating agers and decelerating agers, based on ΔAge. Second, we prioritized proteins with high absolute coefficients in the ProteoAge model, as these proteins contribute most strongly to the BA estimation. Third, we identified proteins whose expression levels showed consistent and significant changes with CA across the cohort. Mortality risk was assessed using Cox proportional hazards models, with AARPs and DARPs evaluated in the same sex cohort from which they were identified—that is, female-derived AARPs/DARPs were analyzed in the female cohort, and male-derived AARPs/DARPs in the male cohort. Specifically, the models were specified as follows:

$${{\rm{h}}}\left(t\right)={h}_{0}\left(t\right)\exp \left({{{\rm{\beta }}}}_{1}\cdot {{\mbox{Protein}}}_{1}+{{{\rm{\beta }}}}_{2}\cdot {{\mbox{Protein}}}_{2}+\cdots +{{{\rm{\beta }}}}_{p}\cdot {{\mbox{Protein}}}_{p}\right)$$

where where h(t) is the hazard at time t, h0(t) is the baseline hazard, protein represents AARPs/DARPs. The models were fit using the coxph() function from the R package survival. HRs and 95% CIs were obtained by exponentiating the regression coefficients. Model outputs were summarized using the tbl_regression() function (with exponentiated results) and the summary() function. The P values derived from the Cox proportional hazards regression models were adjusted for multiple comparisons using the Benjamini-Hochberg (BH) procedure to obtain false discovery rate (FDR) values. Proteins with a statistically significant association with mortality risk (FDR < 0.05) were defined as mortality-aging-associated proteins.

PPI analysis

We employed the STRING database (version 12.0) to evaluate potential PPIs among aging-associated proteins. All associations were derived from high-throughput sequencing data, along with mining from databases and the literature. Cytoscape software (version 3.9.1) is an open source software platform provided by Cytoscape Consortium (4221 Hill St, San Diego, CA 92107), which was used to visualize the PPI network.

Functional enrichment results

GO and KEGG functional enrichment analysis were performed utilizing the enrichGO/enrichKEGG functions of the clusterProfiler R package (version 4.10.1). Categories with FDR < 0.05 were considered significantly enriched, and visualizations were generated using the “dotplot” function.

Statistics and reproducibility

Statistical analysis was implemented using R software (version 4.3.2). The data normalization strategy for the proteomics data of UKB-PPP has been described previously (https://biobank.ndph.ox.ac.uk/showcase/ukb/docs/Olink_1536_B0_to_B7_Normalization.pdf). The Pearson correlation coefficient between the BA and CA was calculated by the sm_statCorr function. The correlation analysis results were visualized using the ggplot2 R package (version 3.5.1). P values were adjusted for multiple comparisons using the FDR via the BH method. A Cox proportional hazards model was constructed using the coxph function of the survival R package (version 3.7–0), and a forest plot was generated with the forest function of the forestploter R package (version 1.1.2).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Characteristics of the cohort

The UKB-PPP collected plasma proteomic profiles over 54,000 participants26,27,36. We included 53,013 individuals with available age information and over 80% proteomics data completeness for subsequent analysis37,38. To improve the generalizability of our study, we analyzed all participants included in the UKB-PPP without imposing additional exclusion criteria.

This cohort comprised 28,579 female participants (aged 39–71 years) and 24,434 male participants (aged 39–71 years). In the female cohort, 25.6% of the individuals (n = 7358) were aged 39–50 years, 35.3% (n = 10,098) were aged 51–60 years, and 38.9% (n = 11,123) were aged 61–70 years. In the male cohort, 25.2% (n = 6167) of the individuals were aged 39–50 years, 32.5% (n = 7940) were aged 51–60 years, and 42.3% (n = 10,327) were aged 61–70 years. The average BMI was 27.2 for females and 27.8 for males. Among the females, 59.2% had never smoked, whereas 48.1% of the males did. At the time of the study, 89.7% of females and 92.8% of males currently consume alcohol. In the female cohort, 40.3% had a history of smoking, with 8.9% being current smokers, whereas in the male cohort, 51.4% had a history of smoking, with 12.5% being current smokers. In the female cohort, 4.1% had diabetes, 25.0% had hypertension, 0.9% had heart disease, and 1.3% had a history of stroke. In the male cohort, 7.6% had diabetes, 32% had hypertension, 4.6% had heart disease, and 2.2% had a history of stroke. Detailed information is provided in Supplementary Data 1 and Supplementary Data 2.

Development of the plasma proteomic aging clock

An overview of this study and workflow is illustrated in Fig. 1. A total of 2923 proteins were detected in the plasma samples utilizing the Olink platform. The entire cohort (53,013 participants from the UKB-PPP database) was divided into a training set (70%, 20,005 in the female cohort and 17,103 in the male cohort) and a validation set (30%, 8574 in the female cohort and 7331 in the male cohort). We conducted a 20-fold cross-validation using the cv.glmnet function in the glmnet R package, with the mean squared error (MSE) employed as the evaluation metric. The lambda.min value was selected as the final λ. During the training phase, we applied elastic net regression to train the proteomic aging clock model, testing 20 values of α at intervals of 0.05 within the range of 0–1. The model with the lowest MSE was identified as the final model. For the female cohort, the model with α = 0.05 was selected, whereas for the male cohort, the model with α = 1 was determined to be optimal (see Supplementary Data 3). After selecting the model with the lowest MSE as the final model, we evaluated the performance of the ProteAge model, a proteomics-based BA prediction model, in the validation set. As shown in Fig. 2, ProteAge was strongly correlated with CA in both females (Fig. 2a, k, R = 0.936, P < 0.001, MAE = 2.225) and males (Fig. 2b, k, R = 0.928, P < 0.001, MAE = 2.424). To avoid the bias caused by the heterogeneity of age distribution, we performed stratified analysis by decade, 40–50, 50–60, and 60–70. Stratified analysis showed the correlation coefficients of ProteAge varied across age group among females, the correlations were 0.759 for the 40–50 age group, 0.758 for the 50–60 group, and 0.718 for the 60–70 group. In males, the corresponding correlations were 0.745, 0.734 and 0.706, respectively (see in Supplementary Fig. 1a–f). Considering the higher correlation of ProteAge constructed from the 40 to 70 age group, this version was selected for subsequent analysis.

Fig. 1: Overview of the study design and primary analysis workflow.
figure 1

a The UKB-PPP database was used for training and testing the ProteAge model. The NHANES III database served as the training set for KDM BA and PhenoAge, whereas the UKB database was used as the validation set. b ProteAge was utilized to predict mortality and identify aging-related proteins.

Full size image
Fig. 2: Evaluation of the performance of the proteomics aging clock.
figure 2

a, b Scatter plots and correlations between ProteAge and CA for the female (R = 0.936, P < 0.001, MAE = 2.225) and male (R = 0.928, P < 0.001, MAE = 2.448) cohorts. c Scatter plots and correlations between KDM BA and CA for females (red, R = 0.904, P < 0.05, MAE = 3.606) and males (blue, R = 0.845, P < 0.05, MAE = 4.169). d Scatter plots and correlations between PhenoAge and CA for females (orange, R = 0.795, P < 0.05, MAE = 7.696) and males (blue, R = 0.822, P < 0.05, MAE = 6.476). Scatter plots and correlations between KDM (R = 0.787, P < 0.001, MAE = 4.714, e)/PhenoAge (R = 0.761, P < 0.001, MAE = 6.371, f) and ProteAge in the male cohort. g, h Scatter plots and correlations between KDM (R = 0.851, P < 0.001, MAE = 3.666, g)/PhenoAge (R = 0.719, P < 0.001, MAE = 7.869, h) and ProteAge in the female cohort. LOESS fitting plots showing the relationship between ProteAge residuals and CA for females (red, i) and males (blue, j). k Bar plot demonstrating the mean absolute error of different aging clock models.

Full size image

Assessing the predictive accuracy of the proteomic aging clock

We also evaluated the performance of two other BA estimation methods (the Klemera-Doubal method (KDM) and PhenoAge). We calculated the correlation between KDM BA and CA in females (R = 0.904, P < 0.001, MAE = 3.666) and males (R = 0.845, P < 0.001, MAE = 4.169, Fig. 2c, k) and the correlation between PhenoAge and CA in females (R = 0.795, P < 0.001, MAE = 7.696) and males (R = 0.822, P < 0.001, MAE = 6.476, Fig. 2d, k). The correlation analysis between KDM BA/PhenoAge and ProteAge indicated that ProteAge was reliable (Fig. 2e–h).

To calculate the deviation between the BA and CA, we regressed CA out of the KDM BA and PhenoAge using a 3-degree-of-freedom natural spline. This produced ΔAge, which represents the difference between the predicted BA and CA. We compared the ΔAge values calculated by the ProteAge, KDM, and PhenoAge models. Detailed data on the BAs predicted by the ProteAge model are presented in Supplementary Data 4. Interestingly, locally estimated scatterplot smoothing (LOESS) analysis revealed distinct aging trajectories between female and male participants. The ΔAge-based aging trajectory of females experienced nonlinear fluctuations at multiple key points, notably around the ages of 48, 58, and 65 (Fig. 2i). whereas males showed a relatively stable trajectory with a single decline between ages 55 and 65 (Fig. 2j). Hormonal changes may be the cause of the fluctuations in female ΔAge across different age groups24. To this end, we analyzed levels of two sex hormones, testosterone (Field ID 30850) and oestradiol (Field ID 30800), along with sex hormone-binding globulin (SHBG, Field ID 30830), estrogen receptor 1 (ESR1), prolactin (PRL), and the sex hormone-regulating proteins, hydroxysteroid 17-beta dehydrogenase 3 (HSD17B3). In males, most hormones and related proteins exhibited relatively stable levels across age. In contrast, females showed pronounced aging-related fluctuations (Supplementary Fig. 2a–c, e–g). Specifically, in female corhort, oestradiol peaked and began declining around age 47, stabilizing after 62 in females. SHBG demonstrated a turning point around age 58, than began increasing. ESR1 expression transitioned from an upward to a downward trend around age 53, with the decline plateauing near age 67. PRL levels rose until approximately age 44, followed by a gradual decline that plateaued around 66. Additionally, HSD17B3 showed a reversal from increasing to decreasing around age 67. To further contextualize these findings, we extracted menopause age data (Field ID 3581) from the UK Biobank. The average age at menopause among the female participants included in our study was 49.53 years (Supplementary Data 2). Differential analysis revealed significant alterations in the proteomic profiles before and after menopaus (Supplementary Fig. 2d). These results highlight menopause as a key biological transition that reshapes the systemic proteomic landscape and may underlie the sex-specific aging patterns observed in ΔAge trajectories.

Increased mortality risk associated with accelerating ProteAge rate

To investigate the biological significance of ProteAge, we assessed its capacity to predict mortality rates using Cox proportional hazards regression analysis. First, we measured the aging rate (ΔAge) by calculating the difference between the predicted BA based on the ProteAge model and CA. Given the limitations inherent in the UKB dataset, we restricted our calculations of HR and 95% CIs to participants aged 40–70 years. In this study, the HR represents the proportional change in mortality risk associated with each one standard-deviation (SD) increment in BA ΔAge. An HR greater than 1 indicates elevated mortality risk, whereas an HR less than 1 suggests a protective association. Additionally, we evaluated the mortality rate prediction capabilities of two alternative BA calculation models, KDM and PhenoAge. The forest plot indicated that ΔAge was associated with mortality among participants in specific age groups. Estimates of the association between one SD increases in ΔAge and the mortality rate are illustrated in Fig. 3. 17,402 participants were excluded due to missing data that prevented the estimation of KDM and PhenoAge-based BAs. As a result, 35,611 participants were included in the final analysis. Among them, 32,105 (90.16%) were right-censored at last contact due to study completion, loss to follow-up, or transfer out. The meidan follow-up time was 9.17 years (range: 0.02–16.58 years, mean 8.62 ± 3.75). During the follow-up period, a total of 5765 death events were observed, of which, 2274 were attributable to cancer. Cancer cases were classified according to ICD-10 codes C00-C97 with histopathological confirmation from medical records and registry linkage. Specifically, among individuals aged 40–70 years, there were 19,156 females (1464 death events) and 16,453 males (2040 death events). For the 40–50 year group, there were 4514 females (87 death events) and 3810 males (99 death events). In the 50–60 year group, we included 6412 females (321 death events) and 5084 males (374 death events). Lastly, in the 60–70 year group, there were 8148 females (1036 death events) and 7483 males (1540 death events). Details are shown in Supplementary Data 2.

Fig. 3: Mortality prediction performance of various aging clock models stratified by age group.
figure 3

Bar plot showing the HRs for all-cause mortality per 1-SD increase in different BA (ΔAge/SD), adjusted for CA. Error bars indicate 95% CI. Panels represent different age groups. a Individuals aged 40–70. b Individuals aged 40-50. c Individuals aged 50-60. d Individuals aged 60–70.

Full size image

Our findings indicated that an increase in ProteAge ΔAge significantly correlated with an increased mortality rate among males across various age groups, including aged 40–70 years (HR = 1.058, 95% CI = 1.015–1.102, P = 0.008), and specifically those aged 60–70 years (HR = 1.049, 95% CI = 1.002-1.099, P = 0.042). These findings suggested that ProteAge ΔAge may serve as a valuable indicator of mortality risk in older male population. Similarly, PhenoAge ΔAge also has demonstrated a strong association with mortality risk in males aged 40–70 (HR = 1.034, 95% CI = 1.001–1.069, P = 0.041) and among those aged 60–70 (HR = 1.053, 95% CI = 1.014–1.094, P = 0.008). Older male population. Meanwhile, KDM showed mortality prediction ability in male population aged 40–70 (HR = 1.042, 95% CI = 1.003–1.083, P = 0.037), with particularly high predictive strength in the younger subgroup aged 40–50 (HR = 1.211, 95% CI = 1.027–1.428, P = 0.023). However, this exceptionally strong association in the 40–50 age group of KDM may reflect statistical bias due to the relatively small number of death events compared to 50–60 age group and 60–70 age group. Detailed results are provided in Supplementary Data 5.

PhenoAge, which was developed using a mortality risk-based scoring of biomarkers scorning system, capturing not only CA information but also biological signals directly linked to mortality, such as inflammatory markers (e.g., C-reactive protein), measures of kidney function (e.g., creatinine), and indicators of nutritional status (e.g., albumin)32,34. This design likey contributes to its strong predictive performance32. Notably, Our proteomics-based model, ProteAge exhibited a mortality prediction accuracy comparable to that of PhenoAge, despite being derived without incorporating mortality-related features, suggesting its potential as an alternative and biologically informative aging biomarker.

Identification and characterization of proteins involved in accelerating BA

Next, we classified participants whose ΔAge was greater than the MAE as “accelerating agers” and those whose ΔAge was less than the MAE as “decelerating agers”. We conducted a differential protein expression analysis of the proteomic profiles between the accelerating and decelerating ager groups using the limma package, with modified guidelines provided by the UKB-RAP (https://github.com/dnanexus/UKB_RAP/tree/main/proteomics/protein_DE_analysis). This analysis employed a threshold |log2(FC)| > log2(1), with a stringent false discovery rate (FDR) < 0.05 to control for multiple testing39,40. This approach allows for the detection of subtle yet potentially meaningful protein changes, particularly for low-abundance regulatory proteins, while the strict FDR filtering ensures that only robust and reproducible differences are prioritized for interpretation. The results of this analysis are presented in a volcano plot for both female (Fig. 4a) and male (Fig. 4b) participants. For the female cohort, a total of 1238 differentially expressed proteins (DEPs) were identified, including 1144 upregulated DEPs and 72 downregulated DEPs (Fig. 4a), and for the male cohort, a total of 753 DEPs, including 609 upregulated DEPs and 144 downregulated DEPs (Fig. 4b), were identified. We noted that PAEP (progestagen-associated endometrial protein) showed a markedly low fold change in the female differential expression analysis (log2FC = −1.32), suggesting that it may act as a negative regulator in the aging process. Some studies have suggested its potential as a biomarker for ectopic pregnancy and certain types of cancer41,42.

Fig. 4: Identification and characteristics of proteins associated with accelerated BA.
figure 4

Volcano plots of DEPs between accelerating agers and decelerating agers in the female (a) and male (b) cohorts. Proteins with an FDR < 0.05 and |log2(fold change)| > log2(1) were considered statistically significant. c, d Venn diagrams of significantly up- and downregulated proteins in the female and male cohorts. Red-labeled proteins have been previously reported in other studies to be closely associated with aging. GO enrichment analysis of upregulated proteins specific to females (e), common to both females and males (f), and specific to males (g). KEGG enrichment analysis of upregulated proteins specific to females (h), common to both females and males (i), and specific to males (j).

Full size image

We identified 506 proteins that were upregulated in both male and female accelerating agers, including NEFL, GFAP and CXCL9, which are widely associated with aging. 660 proteins, such as interleukin 1 receptor antagonist (IL1RN), which functions as an inhibitor of interleukin 1, that were specifically upregulated in female accelerating agers, and 103 proteins, such as TSPAN1and LEFTY2, that were specifically upregulated in male accelerating agers. TSPAN1 is associated with diverse biological functions, such as cell signal transduction, adhesion, migration and protein trafficking43,44(Fig. 4c). Among the proteins downregulated in accelerating agers, 25 proteins were downregulated in male and female accelerating agers (such as CSTV and CA6), 47 were specifically downregulated in female accelerating agers (such as PAEP and CST6), and 119 were specifically downregulated in male accelerating agers (Fig. 4d; see details in Supplementary Data 7). CSTV is a protein associated with the development of various cancers, such as colorectal and breast cancer, and DSG4 is a protein closely associated with cancer and autoimmune diseases. Notably, PAEP showed a markedly low fold change in the female differential expression analysis (log2FC = −1.32), suggesting that it may act as a negative regulator in the aging process. Some studies have suggested its potential as a biomarker for ectopic pregnancy and certain types of cancer41,42.

To further understand the functions of the DEPs associated with accelerating agers and decelerating agers, we subsequently performed Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) and Wikipathway functional enrichment analysis of these proteins (Supplementary Data 8). GO enrichment analysis revealed that the overlapping upregulated DEPs in both males and females were enriched primarily in collagen-containing extracellular matrix (ECM), chemotaxis and taxis (Fig. 4f). In contrast, proteins specifically upregulated in females were enriched predominantly in positive regulation of cell adhesion, leukocyte migration and positive regulation of cytokine production (Fig. 4e). Conversely, proteins specifically upregulated in males were enriched in negative regulation of cell adhesion and ECM-associated pathways (e.g., ECM organization and external encapsulating structure organization, Fig. 4g). The overlapping DEPs observed in both females and males were enriched primarily in cytokine-cytokine receptor interactions and viral protein interactions with cytokines and cytokine receptors (Fig. 4i). DEPs that were specifically upregulated in females were strongly enriched in lipid metabolism, pathways of neurodegeneration-multiple diseases (Fig. 4h), whereas those that were specifically upregulated in males were significantly associated with the JAK-STAT signaling pathway and intestinal immune network for IgA production (Fig. 4j). In addtion, WikiPathways analysis revealed disease-related associations not captured by KEGG. For instance, proteins upregulated in both females and males were enriched in pathways related to pleural mesothelioma, lung fibrosis, spinal cord injury, T cell modulation, and desmoplasia in pancreatic cancer. These findings are consistent with the disease characteristics observed in our cohort and provide valuable complementary insights beyond those identified by KEGG (Supplementary Fig. 3a–c).

Extracting proteins with non-zero coefficients in the ProteAge model

In the ProteAge model, proteins with non-zero coefficients are retained, meaning that they contribute to the ProteAge aging rate. If the coefficient of one protein coefficient is exactly zero, that protein is excluded from the ProteAge model. To determine the proteins positively contributing to the aging rate and those negatively contributing to the aging rate, we extracted the coefficients of all 2923 proteins in the ProteAge model (Fig. 5a, Supplementary Data 9) and visualized the top 20 proteins with the highest/lowest coefficients (Fig. 5b, c). For instance, ROBO1 exhibited high negative coefficients, while ITGAV, SCARF2 and LTBP2 showed high positive coefficients in both male and female, suggesting their respective roles in decelerating and accelerating the proteomic aging rate.

Fig. 5: Extraction and function analysis of non-zero coefficients from the ProteAge model.
figure 5

a Workflow for the development of the ProteAge model. Identification of proteins with non-zero coefficients in the female (b) and male cohorts (c). Blue proteins indicate those with consistently lower coefficients in both male and female cohort models, while red proteins indicate those with consistently higher coefficients in both models. Both blue and red proteins have been previously reported to be associated with aging. GSEA of proteins with non-zero coefficients in the female (d) and male (e) cohorts.

Full size image

Next, we performed GSEA on proteins with non-zero coefficients on the basis of the GO database. GSEA analysis revealed that the difference lines in the focus on vascular and cellular signaling in females, but vascular development, oxidative stress and structure maintenance in male. Both male and female share the regulation of kinase activity and cell migration, suggesting some common molecular mechanisms despite sex-specific variations (Fig. 5d, e).

Identification of proteins strongly associated with the aging process

After the identification of DEPs between accelerating agers and decelerating agers, the coefficients of all 2923 proteins in ProteAge were extracted, which helped us identify aging-associated proteins. DEPs facilitated our understanding of stage-specific alterations in protein levels during aging, whereas proteins with non-zero coefficients offered insights into those functions as accelerating/decelerating aging rate factors. As described in the review published in 2023, the time-dependent manifestation of alterations accompanying the aging process must be the criterion of each hallmark of aging7. Therefore, we next investigated the persistent changes in protein expression throughout the aging process. Additionally, we mapped the expression profiles of all 2923 proteins during aging. To this end, we calculated the mean values of all 2923 proteins at each age stage and identified aging-associated changes in these proteins utilizing the “Mfuzz” clustering method. Clustering analysis was conducted separately for males and females, with the number of clusters fixed at five.

In females, proteins in cluster 2 (C2, n = 1,430), such as LTBP2, EDA2R, NEFL, and GDF15, presented continuously upregulated expression patterns. Conversely, proteins in cluster 5 (C5, n = 404), including KIT and CSTV (also known as cathepsin v), continuously decreased in expression with age (Fig. 6a, Supplementary Data 10). Similarly, the clustering analysis for the males revealed a cluster of proteins (C2, n = 1034) whose expression levels continuously increased with age and a cluster of proteins (C5, n = 492) whose expression levels continuously decreased with age (Fig. 6f, Supplementary Data 10).

Fig. 6: Identification and charaterization of age-dependent proteins.
figure 6

a Heatmap showing the clustering results of proteins whose expression changed with age in the female cohort. b Heatmap depicting the clustering results of proteins whose expression changed with age in the male cohort. c KEGG analysis results for proteins in the C2 cluster in females. d KEGG analysis results of proteins in the C5 cluster in females. e KEGG analysis results of proteins in the C2 cluster in males. f KEGG analysis results of proteins in the C5 cluster in males.

Full size image

We subsequently performed GO and KEGG enrichment analysis on the C2 and C5 proteins in both males and females (Supplementary Data 11). The results of the KEGG analysis of proteins whose expression changed continuously with age are shown in Fig. 6c–f. The results revealed significant enrichment of C2 cluster proteins in immune-related pathways in both the male and female cohorts. Notably, in females, C2 cluster proteins were specifically enriched in the focal adhesion pathways. In contrast, in males, C2 cluster proteins were selectively enriched in the JAK-STAT signaling pathway, rheumatoid arthritis pathway, and Ras signaling pathway. For C5 cluster proteins, KEGG analysis revealed significant enrichment in the PI3K-Akt signaling pathway, lipid metabolism, immune pathways in both gender. In addition, In females, C5 cluster proteins were significantly enriched in neuroactive ligand-receptor interaction, cAMP signaling pathway, pathways of neurodegeneration-multiple diseases, Ras signaling pathway. In males, C5 cluster proteins were significantly enriched in the cell adhesion molecules, human papillomavirus infection. (Supplementary Fig. 5c–f).

Integrated analysis revealed accelerating/decelerating ARPs

To identify proteins involved in accelerating/decelerating aging rates, we employed three distinct strategies: differential expression analysis between accelerating and decelerating agers, identification of proteins exhibiting continuous changes with aging, and extraction of proteins with non-zero coefficients in the ProteAge model. We defined the intersection of proteins screened by these three approaches as AARPs and DARPs. As illustrated in the Venn diagram, there were 414 AARPs in females (Fig. 7a), which was more than the 248 AARPs in males (Fig. 7e). As shown in Fig. 2i, j, the aging trajectory in females exhibited greater fluctuations than in males. In addition, a pronounced shift in the proteomic profile was observed before and after menopause in females (Supplementary Fig. 2d). Together, these findings may partially explain why a greater number of AARPs were identified in females compared to males in this study. The numbers of DARPs identified in females and males were comparable: 40 DARPs in females (Fig. 7c) and 75 DARPs (Fig. 7g) in males. All identified proteins are listed in Supplementary Data 12. Moreover, we observed 148 overlapping AARPs in both females and males and 14 overlapping DARPs in both females and males (Fig. 7i, k). These overlapping AARPs and DARPs represent similar protein-level molecular changes during the aging process in both males and females. We also checked the expression levels of these 148 overlapping AARPs and 14 overlapping DARPs in both females and males (Supplementary Fig. 7).

Fig. 7: Identification and characterization of aging-associated proteins (AARPs and DARPs) via intersection analysis.
figure 7

a Venn diagrams showing the AARPs in females. b PPI network of AARPs in females (the top 20 proteins with the highest degree). The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. c Venn diagrams showing the DARPs in females. d PPI network of DARPs in females. The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. e Venn diagrams showing the AARPs in males. f PPI network of AARPs in males (the top 20 proteins with the highest degree). The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. g Venn diagrams showing the DARPs in males. h PPI network of DARPs in males (the top 20 proteins with the highest degree). The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. i Venn diagrams showing the overlapping AARPs in females and males. j PPI network of AARPs in females and males (the top 20 proteins with the highest degree). The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. k Venn diagrams showing the overlapping DARPs in females and males. l PPI network of DARPs in females and males. The size of each node represents the degree of connectivity of the corresponding protein in the PPI network, with larger nodes indicating higher connectivity. Bar plots depicting the results of the GO (m) and KEGG (n) enrichment analysis of AARPs in both females and males.

Full size image

We then constructed protein-protein interaction (PPI) networks for the AARPs and DARPs in both females and males. Using the STRING database to predict PPI, we analyzed the resulting networks in Cytoscape by examining their topological properties through the “Analyze Network” function. We then identified and highlighted the accelerating aging rate protein (AARPs) and DARPs with the highest degrees of connectivity. The PPI network analysis of AARPs in females revealed several key proteins with high connectivity, including IL6, FN1, and CD4 (Fig. 7b). CD8A, KIT, SHC1, and PRL, and emerged as the most significantly connected DARPs (Fig. 7d). For males, TGFB1, CCL4, CXCL10, and CXCL9 were AARPs with highest connection degree, CD8A, EGFR, PLG, and KIT were DARPs with highest connection degree (Fig. 7f, h). An integrated analysis of the male and female cohorts revealed a set of proteins, such as DCN, CCL4, CDH2, and TREM2, which were identified as AARPs in both females and males (Fig. 7j). Meanwhile, 14 proteins (including CD8A, RET, and KIT) were identified as DARPs in both females and males.

Among the AARPs identified in both males and females, CXCL9, a chemokine from the CXC family, were both linked to GO terms related to chemotaxis, taxis, and cytokine activity (Fig. 7m, Supplementary Data 12). CXCL9, also known as MIG (monokine induced by gamma interferon), is a pro-inflammatory chemokine that primarily recruits activated T cells through the CXCR3 receptor and plays a key role in immune surveillance. Its expression tends to increase with age, contributing to the heightened pro-inflammatory state associated with immunosenescence and aging-related tissue damage45. TREM2, a transmembrane receptor primarily expressed on immune cells such as monocytes and macrophages, plays a wide-ranging role in the nervous system; Mutations in TREM2 have been reported to elevate the risk of neurodegenerative disorders, including Alzheimer’s disease, in line with previous findings on the importance of TREM2 in brain health46.

Next, we performed GO and KEGG enrichment analysis on these aging-associated proteins identified in both males and females. GO analysis indicated that the majority of AARPs were significantly enriched in immune-related pathways (e.g., chemotaxis and cell chemotaxis) and ECM-related pathways (Fig. 7m). KEGG analysis indicated that the majority of AARPs were significantly enriched in several pathways, including cytokine-cytokine receptor interactions and chemokine signaling pathways (Fig. 7n). Similar results were observed via GSEA (Supplementary Fig. 4). All the GO/KEGG enrichment analysis results are shown in Supplementary Data 12.

Identifying mortality-aging associated proteins

After identifying AARPs and DARPs based on three criteria, we conducted sex-specific Cox regression analysis to evaluate their associations with mortality. Importantly, AARPs identified in the male cohort were exclusively analyzed for mortality prediction in males, and DARPs identified in the female cohort were exclusively analyzed in females, ensuring that the predictive evaluation was strictly sex-matched and independent. We then extracted the HR, 95% CI, and P value for each protein from the Cox proportional hazaids models. The P values were subsequently adjusted for multiple comparisons using the BH method. Proteins with FDR < 0.05 in the Cox models were defined as mortality-associated aging-related proteins (Fig. 8a). As a result, a total of 172 and 1 mortality-aging-associated proteins were identified in males and females, respectively, all details are shown in Supplementary Data 13. The top 15 proteins with the highest HRs and the bottom 5 proteins with the lowest HRs were extracted from the male and female cohorts, respectively, and visualized in forest plots. AARPs/DARPs with an FDR < 0.05 were highlighted in red or blue and shown in bold (Fig. 8b, c).

Fig. 8: Identification of sex-specific mortality-aging associated proteins.
figure 8

aThe workflow employed to assess the mortality predictive performance of AARPs/DARPs. b Mortality prediction performance of accelerating/decelerating ARPs in females. The boxes represent point estimation, while the horizontal lines indicate the 95% CI. Proteins with a P ≥ 0.05 are displayed in black font, and mortality-aging-associated proteins are shown in bold text. c Mortality prediction performance of accelerating/decelerating ARPs in males. The boxes represent point estimation, while the horizontal lines indicate the 95% CI. Proteins with a P ≥ 0.05 are displayed in black font, and mortality-aging-associated proteins are shown in bold text.

Full size image

In females, ZNRD2 were identified the only mortality-aging-associated protein (HR = 1.130, 95% CI = 1.070–1.200, FDR = 0.018). In males, a total of 172 proteins, the majority of AARPs exhibited HRs greater than 0 in the mortality analysis, indicating an increased risk of death, most DARPs showed HRs less than 0, suggesting a protective effect. For example, of the AARPs, OSMR (HR = 1.550, 95% CI = 1.330–1.800, FDR < 0.001), LTBP2 (HR = 1.340, 95% CI = 1.240–1.450, FDR < 0.001), NRP1 (HR = 1.320, 95% CI = 1.190–1.470, FDR < 0.001), NRP2 (HR = 1.320, 95% CI = 1.190–1.460, FDR < 0.001), and GDF15 (HR = 1.230, 95% CI = 1.170–1.290, FDR < 0.001), were identified as mortality-aging-associated proteins. GDF15 plays an important role in modulating inflammation, metabolic homeostasis, and cellular stress responses is closely linked to aging and contributes to the pathogenesis of various aging-related chronic diseases, including neurodegenerative and cardiovascular diseases29,47,48. Notably, although ITGAV was identified as an AARP, it was associated with a reduced risk of mortality in the Cox analysis. ITGAV (Integrin alpha V) is a member of the integrin family of cell surface receptors that mediate cell adhesion and signaling. It plays a critical role in various biological processes, including angiogenesis, inflammation, and tissue remodeling, and has been implicated in aging and aging-related diseases49,50.

Discussion

Our study presents ProteAge, a proteomic-based aging clock, developed from 53,013 plasma samples from the UKB-PPP. This model was specifically designed to assess BA using elastic net regression. Compared to existing models such as KDM and PhenoAge, ProteAge demonstrated a stronger correlation with CA and lower MAE, indicating its potential as a more precise aging clock. However, these results should be interpreted cautiously, as ProteAge has not yet undergone cross-cohort validation. Both KDM and PhenoAge have been validated through cross-cohort datasets, utilizing the NHANES dataset as the training set and the UKB dataset as the validation set. Therefore, further validation of ProteAge across diverse population is necessary to ensure its reliability.

One of the primary goals of aging research is to develop a robust metric that accurately reflects overall health status and predicts health outcomes. The poor overall health status is associated with an increased risk of adverse events, including suffering chronic disease and mortality. In this study, we demonstrated that ProteAge exhibited promising predictive capacity for mortality. Specifically, Cox proportional hazard analysis showed that ProteAge ΔAge predicted mortality risk in male cohorts, aged 40–70 years, and aged 60–70 years as well, showing a prediction accuracy comparable to that of PhenoAge. Thus, these findings suggest ProteAge not only serves as a reliable metric for assessing BA but also offers a valuable indicator of overall health status, particularly in older males.

Importantly, our study revealed distinct aging characteristics between males and females throughout the aging process. First of all, aging trajectories based on ΔAge from ProteAge, KDM, and PhenoAge revealed more pronounced nonlinear patterns in females than in males. These nonlinear trajectories were particularly evident around menopause, reflecting the complex and dynamic nature of female aging. Secondly, ProteAge ΔAge significantly predicted mortality, particularly in males aged 40–70 and 60–70 age group, but showed weaker performance in females. This may reflect the more complex, nonlinear aging patterns in females—especially around menopause—which challenge the Cox model’s predictive accuracy. Thirdly, we identified sex-specific markers that are robustly associated with the aging process, namely aging-accelerating rate proteins (AARPs) and DARPs. Males and females displayed distinct aging-associated protein landscapes as revealed by AARPs and DARPs, underscoring the biological differences between sexes in the context of aging. These results highlight the need to account for sex-specific biological differences in aging clock development.

The identification of sex-specific patterns and markers prompted us to investigate whether these findings align with the most recent updates to the biological aging framework. In 2025, Kroemer et al. expanded the twelve Hallmarks of Aging to fourteen by adding ECM changes and psychosocial isolation51. Our functional enrichment analysis of 166 aging-associated proteins (AARPs), which were both identified in males and females, revealed significant overrepresentation of pathways related to the ECM and chronic inflammation, closely aligning with two updated hallmarks. Among the identified AARPs, CXCL9, was identified as AARPs in both males and females, exhibited high connectivity in the PPI analysis. CXCL9 is associated with cardiovascular decline during aging, and silencing it can reverse its harmful effects on aging endothelial cells45. In 2024, Da Hea Seo et al. reported a significant association between CXCL9 and physical function, as well as changes in bone mineral density and osteoporotic fracture risk52. Additionally, in the male cohort, GDF15 was identified as an aging-mortality associated protein. GDF15, originally characterized as macrophage-inhibiting cytokine-1, belongs to the TGF-β superfamily53. GDF15 expression is low in healthy young individuals54,55,56,57, but increases significantly in aging-related conditions such as cardiovascular diseases, type 2 diabetes, neurodegenerative disorders, and cancer58,59. The mechanistic link between these markers and the newly recognized hallmarks is further supported by our observation of significant enrichment in the MAPK cascade, a key pathway implicated in aging processes such as genomic instability60,61, telomere attrition62, epigenetic alterations63, and cellular senescence64. MAPK pathways also contribute to inflammation through cytokine modulation65,66,67. Therefore, although Olink platform’s limited coverage (2923 proteins) may have constrained our analysis, our findings highlight the potential of plasma proteomics to identify aging biomarkers across diverse biological systems, providing valuable insights into the molecular mechanisms of aging.

This study is based on UK Biobank data, primarily from individuals of European ancestry, which may limit generalizability to other ethnic or geographic populations. Additionally, the cohort is skewed toward middle-aged and older adults, with limited data from younger individuals, restricting insights into proteomic aging across the full lifespan, especially early development. Moreover, our analysis focused solely on protein expression, which may overlook key aspects of aging-related proteomic complexity. As Qian Zhong et al. mentioned68, post-translational modifications-e.g., phosphorylation, acetylation, glycosylation—likely contribute to aging and could inform complementary aging clocks. Additionally, inspired by Kroemer et al.51, who emphasize genomic instability as a hallmark of aging, point mutation-based clocks may offer novel insights.

Conclusion

ProteAge is a proteomic aging clock that outperforms existing models in estimating BA and predicting mortality, particularly in older males. It reveals sex-specific aging patterns and aligns with updated aging hallmarks such as ECM changes and inflammation. While further validation across diverse populations is needed, our findings underscore the potential of plasma proteomics in aging research and highlight the importance of incorporating sex differences and molecular mechanisms into aging clock development.