Glycemic status-dependent proteomic signatures of biological aging for health risk prediction

Abstract

Existing proteomic aging clocks have been derived from the overall population, with little consideration of extended models tailored to individuals with different glycemic status. We aimed to quantify glycemic status-dependent proteomic signatures of aging and developed proteomic aging scores (ProAS) for health risk prediction. A total of 2923 plasma proteins were measured using Olink in 46,047 UK Biobank participants, including 37,353 with normoglycemia, 5977 with prediabetes, and 2717 with diabetes. Using a three-step screening approach, we identified 11, 23, and 21 representative protein biomarkers associated with all-cause mortality among individuals with normoglycemia, prediabetes, and diabetes, respectively. Three proteins (GDF15, EDA2R, and WFDC2) were shared across all groups, with GDF15 emerging as the top-ranked important protein in normoglycemia and prediabetes and WFDC2 in diabetes. The protein-based ProAS according to glycemic status showed significant associations with diverse health outcomes. Adding the ProAS in the models improved the predictive accuracy of mortality and incident diseases beyond conventional risk factors, but the performance progressively diminished as glycemic status deteriorated. In addition, 72, 51, and 36 out of 102 modifiable factors spanning seven categories were identified as determinants for ProRS in normoglycemia, prediabetes, and diabetes, respectively. Our findings extend the current proteomic clocks by revealing glycemic status-specific aging patterns and their ability to predict age-related outcomes, potentially refining risk stratification and targeted interventions for healthy aging.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11357-025-01950-w.

Introduction

Aging is a complex process characterized by a progressive loss of physiological integrity and function, ultimately triggering the initiation and progression of diseases and death [1, 2]. Chronological age is a strong but imperfect surrogate for aging measures as it could not fully capture the heterogeneous aging rates across individuals that engender differential susceptibilities to disease and premature death. To more precisely measure biological aging, various omics-based aging clocks have been developed [3]. These include first-generation aging clocks, which aim at tracing chronological age [4, 5], and second-generation clocks, which are designed to predict aging-related outcomes (e.g., mortality) [6, 7].

Loss of proteostasis is regarded as a primary hallmark of aging [1]. Blood proteomic profiling, thus, offers a novel avenue for dissecting biological aging processes and uncovering therapeutic targets for age-related disorders [8–10]. Several previous studies have identified proteins associated with aging indicators (chronological age or mortality) and used them to develop proteomic clocks to predict aging-related diseases and mortality [11–15]. However, all of these proteomic clocks have been derived from the whole population, and their applicability to people with abnormal glycemic status has not been systematically examined. Previous studies have found that the aging rate varies dramatically among individuals with different glycemic status [16–18]. In a gender- and age-matched study, participants with prediabetes and diabetes showed an average increase in biological age, which was estimated via routine clinical biomarkers, of 2.69 years and 12.02 years compared to those with normoglycemia [17], respectively, suggesting a more pronounced acceleration of biological aging with worsening glycemic status. Meanwhile, the significantly increased risks of aging-related diseases such as cardiovascular disease (CVD) and chronic kidney disease (CKD) among individuals with prediabetes and diabetes may also indicate the heterogeneity in aging across glycemic status. Given the escalating burden of diabetes and prediabetes globally, efforts to quantify glycemic status-specific biological aging would have significant consequences for developing personalized preventive programs and interventions, thus promoting healthy aging and longevity, especially among vulnerable populations.

In this study, we applied a data-driven approach based on large-scale blood proteomic data from the UK Biobank to develop glycemic status-specific proteomic aging scores (ProAS). Next, we assessed the performance of ProAS by examining their associations with mortality and common age-related diseases, and comparing their predictive ability to several conventional aging metrics. Finally, we conducted an exposome-wide association analysis to systematically identify the modifiable factors for ProAS.

Methods

Study participants

The UK Biobank is a prospective cohort study of over 50,0000 participants aged 40–69 years who were recruited from 22 assessment centers across the UK between 2006 and 2010 [19]. The enrollment procedure encompassed completion of a touch-screen questionnaire, physical measurements, and biological sample collection. Ethical approval was obtained from the North West Multi-Center Research Ethics Committee (11/NW/0382), and all participants provided signed informed consent. This study restricted the sample to 53,020 participants with proteomic profiling data available at baseline who were randomly sampled from the UK Biobank population [20]. These participants selected for proteomic profiling have been shown to be highly comparable to the overall UK Biobank population [20]. After excluding individuals diagnosed with type 1 diabetes and those with over 50% missing data in proteomics measurements [13, 21], the final analytical cohort comprised 46,047 participants (Figure S1). The overall study design is illustrated in Graphical abstract.

Assessment of glycemic status

Participants with type 2 diabetes at baseline were identified through integration of multiple data sources. The UK Biobank algorithms for identification of prevalent type 2 diabetes based on self-reported data, including medical history and medication information, have been described [22]. Admissions and diagnoses data from hospital inpatient records were used to ascertain type 2 diabetes with the International Classification of Disease (ICD)−10 code E11. Diabetes coded as E14 (unspecified diabetes) was also assigned as type 2 diabetes, as the UK Biobank included only middle-aged and elderly participants, for whom unspecified diabetes was primarily type 2 diabetes [23–25]. Undiagnosed diabetes was defined as a glycated hemoglobin (HbA1c) level of ≥ 48.0 mmol/mol (≥ 6.5%) according to the American Diabetes Association criteria [26]. In participants without diabetes, prediabetes was defined as an HbA1c level of 39.0–47.0 mmol/mol (5.7%–6.4%), and normoglycemia was defined as an HbA1c level of < 39.0 mmol/mol (< 5.7%).

Proteomic profiling

The proteomic profiling for plasma samples from UK Biobank participants was undertaken via the Olink Explore 3072 platform, which covered 2923 unique proteins spanning four panels, including cardiometabolic, inflammation, neurology, and oncology. Under stringent quality control procedures, the protein measurements were provided as Normalized Protein eXpression (NPX) values (https://biobank.ndph.ox.ac.uk/showcase/ukb/docs/Olink_1536_B0_to_B7_Normalization.pdf)[27]. Full details on sample selection, processing, and quality control have been described elsewhere [20, 28]. Four proteins with missing measurements in over 20% of the samples were excluded, leaving a total of 2919 proteins for analyses (Table S1). For the remaining proteins with missing values, the K-nearest neighbors (KNN) method was used for imputation (K = 10). Protein levels were rank-based inverse normalized and scaled to have a mean of 0 and standard deviation (SD) of 1 before analyses.

Ascertainment of health outcomes

Information on date and cause of death was obtained through the linkage to national death registries. The all-cause and cause-specific mortality (CVD and cancer) were ascertained using ICD-10 codes (Table S2). Follow-up time was calculated from the date of baseline assessment to the date of death or end of follow-up (Dec 31, 2023), whichever came first. Incident diagnoses of aging-related diseases (composite CVD, coronary heart disease [CHD], stroke, heart failure [HF], atrial fibrillation [AF], CKD, peripheral artery disease [PAD], dementia, and cancer) were ascertained through linkage to hospital inpatient records, and cancer and death registry data using ICD-10 codes (Table S2). Follow-up time was calculated from the date of baseline assessment to the date of first diagnosis of the disease, death, or end of follow-up (Dec 31, 2023), whichever came first.

Development of glycemic status-specific proteomic aging scores (ProAS)

Following glycemic status stratification, the dataset was randomly split into a training and a testing set at a ratio of 6:4 [29, 30], to ensure sufficient outcome events within each glycemic status group. We undertook a three-step approach to identify the optimal protein panel associated with biological aging in the training sets. First, Cox proportional hazard regression models were used to estimate the association between each protein biomarker and all-cause mortality, with adjustment for covariates. All-cause mortality was chosen as a surrogate endpoint, as it represents the ultimate consequence of progressive physiological decline and more robustly captures biological aging signals [31–33]. Bonferroni corrections were applied (P < 0.05/2919). For significant proteins, KEGG and GO-BP enrichment analyses were conducted to examine the pathways and biological processes related to the target genes for proteins using the online platform Enrichr (https://maayanlab.cloud/Enrichr/)[34].

Second, a Cox regression model with least absolute shrinkage and selection operator (LASSO) penalization was trained on plasma proteins that showed a significant association with all-cause mortality in the training set. We tuned the hyperparameter lambda through tenfold cross-validation, where the chosen lambda yielded a mean squared error within one standard error of the minimum. Proteins with nonzero coefficients were retained for further selection.

Third, we employed a machine learning-based approach to identify key biomarkers of aging, combining rigorous cross-validation with an interpretable Shapley additive explanations (SHAP) model to minimize spurious associations. Specifically, the proteins selected by LASSO were fed into a preliminary trained light gradient boosting machine (LightGBM) classifier using fivefold cross-validation [35], and each protein was ranked given its contribution to the predictive performance of the model (as judged by the information gain). Then, we selected the important proteins according to a threshold of mean information gain value ≥ 0.02 and visualized these proteins using SHAP plots.

Based on the proteins selected above, we constructed glycemic status-specific ProAS, using the LightGBM model for predicting all-cause mortality across different glycemic status. To reduce overfitting, the model was trained via fivefold cross-validation in the training set, with hyperparameters optimized using Bayesian optimization to maximize the area under the curve (AUC) (see Table S3 for all possible hyperparameter combinations). Once the hyperparameters were determined, the LightGBM classifier was then retrained for all individuals in the training set.

Calculation of different aging metrics

To compare the predictive performance of established ProAS with other aging indicators, two conventional aging metrics were utilized: frailty index (FI) and telomere length (TL). Detailed information on assessment of aging metrics is shown in Supplementary methods.

Modifiable factors

A total of 102 potentially modifiable factors assessed at the UK Biobank baseline survey were included to test their associations with ProAS. These factors were generally classified into seven categories: local environment (e.g., air pollution), psychosocial factors (e.g., irritability), socioeconomic status (e.g., Townsend deprivation index [TDI]), medical history (e.g., prevalent CVD at baseline), early life factors (e.g., breastfed as a baby), physical measures (e.g., body fat percentage), and lifestyle (e.g., diet score). Details of the factor processing are presented in Table S4.

Assessment of covariates

Information on covariates collection was provided in Supplementary Methods.

Statistical analyses

Baseline characteristics were described as mean (SD) and count (percentage). The differences across different glycemic status groups were assessed using analysis of variance (ANOVA) for continuous variables and χ [2] test for categorical variables. Missing data for covariates were imputed using the KNN algorithm for continuous variables and the mode for categorical variables.

The correlations of ProAS with FI, TL, and chronological age in the testing set were estimated using Pearson’s correlation. Then, we applied Cox regression models for all-cause mortality and Fine-Gray subdistribution hazard models for cause-specific mortality and incident diseases to test the associations of ProAS with mortality and aging-related diseases. Models were adjusted for chronological age, sex, race, TDI, body mass index [BMI], never smoking, healthy drinking, physical activity, diet score, sleep score, CVD, cancer, high cholesterol, hypertension, respiratory disease, and antihypertensive and cholesterol-lowering medications. Among participants with diabetes, the model was additionally adjusted for diabetes duration, HbA1c, and diabetes medication use.

Next, time-dependent receiver operating characteristic (ROC) curves (5, 10, and 15 years) were performed to evaluate the predictive performance of ProAS and other aging measures for all-cause mortality across different glycemic status. Participants were then categorized into three groups according to the tertiles of ProAS. To validate the predictive performance independent of chronological age, we calculated the residuals of ProAS, FI, and TL by regressing them against chronological age and repeated the ROC analyses. In addition, we conducted age-stratified analyses to explore the predictive performance of ProAS across chronological age groups (≤ 60 and > 60 years).

We further investigated the predictive value of different aging metrics for aging-related disease incidence within 10 years. First, AUC was used to compare predictive utility between ProAS, FI, TL, and chronological age independently. Then, we assessed the added predictive value of incorporating ProAS into disease prediction models with increasingly incremental sets of traditional risk factors. Demographics-adjusted models included age and sex. Full conventional risk factors models were further adjusted for common risk factors. DeLong’s test was used to assess the differences in 10-year AUC estimates resulting from the addition of the ProAS beyond each set of covariates. Furthermore, the 10-year continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were used to capture finer increments in reclassification.

Modifiable factors for ProAS

Multivariable linear regression models were used to examine the associations between individual modifiable factors and ProAS, with a Bonferroni-corrected significance threshold (P < 0.05/102 = 4.90 × 10⁻ [4]). Continuous modifiable variables were Z-normalized, and results are presented as β coefficients per 1 − SD increment in the corresponding factor. Additionally, continuous factors were categorized into tertiles, with the lowest tertile as the reference. The models were adjusted for age, sex, race, and the use of antihypertensive and cholesterol-lowering medications, as well as diabetes medication use for participants with diabetes only.

Relevant analyses were conducted using Python (v.3.12.4) and R (v.4.4.1). We considered two-tailed P < 0.05 to be significant.

Results

Population characteristics

Among the 46,047 participants with complete data on proteomic profiling, the mean age was 56.8 years (SD, 8.2) and 54.1% were female. Of these participants, 37,353 were classified as normoglycemia, 5,977 as prediabetes, and 2,717 as diabetes. Characteristics of participants according to glycemic status are shown in Table 1. Compared with participants with normoglycemia, those with prediabetes or diabetes were more likely to be older, male, more deprived, and had higher BMI, less physical activity, a lower healthy diet score and sleep score, a higher prevalence of chronic diseases, and greater use of blood pressure- and cholesterol-lowering medications.

Table 1.

Baseline characteristics of UK Biobank participants by glycemic status

Characteristics	Overall (n = 46,047)	Normoglycemia (n = 37,353)	Prediabetes (n = 5977)	Diabetes (n = 2717)
Age, years	56.8 (8.2)	56.1 (8.3)	59.7 (7.2)	60.0 (7.1)
Female	24,920 (54.1)	20,616 (55.2)	3279 (54.9)	1025 (37.7)
White ethnicity	42,907 (93.2)	35,311 (94.5)	5310 (88.8)	2286 (84.1)
Townsend deprivation index	− 1.1 (3.3)	− 1.3 (3.2)	− 0.8 (3.4)	− 0.1 (3.7)
BMI, kg/m2	27.4 (4.8)	26.9 (4.5)	28.8 (5.2)	31.4 (5.9)
Never smoking	25,097 (54.5)	20,941 (56.1)	2916 (48.8)	1240 (45.6)
Healthy drinking	30,301 (65.8)	23,921 (64.0)	4273 (71.5)	2107 (77.5)
Diet score	3.1 (1.2)	3.1 (1.2)	3.0 (1.2)	3.0 (1.2)
Physically active	29,010 (63.0)	23,743 (63.6)	3729 (62.4)	1538 (56.6)
Sleep score	3.7 (0.9)	3.7 (0.9)	3.6 (1.0)	3.5 (1.0)
Cardiovascular disease	4014 (8.7)	2493 (6.7)	851 (14.2)	670 (24.7)
Cancer	4428 (9.6)	3463 (9.3)	670 (11.2)	295 (10.9)
High cholesterol	7189 (15.6)	4522 (12.1)	1416 (23.7)	1251 (46.0)
Hypertension	12,786 (27.8)	8845 (23.7)	2208 (36.9)	1733 (63.8)
Respiratory disease	1889 (4.1)	1328 (3.6)	379 (6.3)	182 (6.7)
SBP, mmHg	137.7 (19.0)	136.9 (18.9)	140.9 (19.4)	141.3 (18.7)
Total cholesterol, mg/dL	219.2 (43.6)	221.6 (41.6)	221.5 (46.9)	180.0 (44.7)
HbA1c, mmol/mol	5.5 (0.6)	5.3 (0.3)	5.9 (0.2)	6.9 (1.3)
Creatinine, mg/dL	0.8 (0.2)	0.8 (0.2)	0.8 (0.2)	0.9 (0.2)
Antihypertensive medication	10,001 (21.7)	6499 (17.4)	1876 (31.4)	1626 (59.8)
Cholesterol-lowering medication	8381 (18.2)	4834 (12.9)	1663 (27.8)	1884 (69.3)

Data are presented mean (SD) or n (%). P values were calculated using analysis of variance and χ2 test for continuous and categorical variables, respectively. All P values were < 0.001

BMI body mass index, HbA1c hemoglobin A1c, SBP systolic blood pressure

Development of glycemic status-specific ProAS

To develop the ProAS, the cohort was randomly split into training and testing sets with a 6 to 4 ratio. In the training set, 577, 302, and 333 out of the 2919 proteins exhibited significant associations with all-cause mortality (P < 0.05/2919) among participants with normoglycemia, prediabetes, and diabetes, respectively (Figure S2 and Table S5). GDF15 and WFDC2 demonstrated the most significant association with all-cause mortality in normoglycemia and prediabetes. For diabetes, WFDC2 and EDA2R had the most significant association with all-cause mortality. Using KEGG methods, the significant proteins were mainly enriched into pathways related to immune function and cell proliferation, differentiation, and survival (Figure S2). Furthermore, GO-BP enrichment analyses showed that these proteins were mainly involved in cytokine-mediated signaling pathway and inflammatory response. Specific enriched pathways for the significant proteins are presented in Table S6 and Table S7.

The LASSO Cox model with tenfold cross-validation further identified a combination of 63, 30, and 28 proteomic biomarkers for mortality prediction in individuals with normoglycemia, prediabetes, and diabetes, respectively (Table S8). Utilizing a threshold of mean information gain value ≥ 0.02 with the LightGBM model, 11, 23, and 21 proteins were finally selected for the development of ProAS in individuals with normoglycemia, prediabetes, and diabetes, respectively (Fig. 1 and Table S9). Among them, GDF15, NEFL, EDAR2, and WFDC2 were the top four important proteins related to all-cause mortality in participants with normoglycemia and prediabetes. In contrast, in participants with diabetes, WFDC2 stood out as the top-ranked predictive protein, followed by EDAR2, NT-proBNP, and CLEC3B. Of particular interest was the fact that several proteins (i.e., WFDC2, NT-proBNP, and CLEC3B) played an increasingly important role in predicting mortality with the worsening of glycemic status.

Fig. 1.

Identification of plasma proteins for all-cause mortality by LightGBM model by glycemic status. A Ranking of plasma protein importance. The bar chart indicates the importance of the sorted proteins based on their contributions to the prediction of all-cause mortality (as judged by the information gain). B SHAP visualization plot of selected proteins. SHAP Shapley additive explanations

Next, glycemic status-specific ProAS were developed using the LightGBM model. ProAS were highly correlated with chronological age in the testing set, and the coefficient of correlation was gradually reduced as the glycemic status deteriorated. Moreover, ProAS showed a moderate correlation with FI and TL (Figure S3). Of note, males exhibited significantly higher ProAS than females among participants with normoglycemia and prediabetes, suggesting an elevated overall mortality risk for males (Figure S4).

Associations of glycemic status-specific ProAS with mortality

The associations of glycemic status-specific ProAS with all-cause and cause-specific mortality in the testing set are shown in Fig. 2. After adjustment for potential confounders, each 1 SD increase in ProAS was associated with a 140% (HR: 2.40, 95% CI: 2.23–2.58), 122% (2.22, 1.94–2.53), and 121% (2.21, 1.89–2.57) higher risk of all-cause mortality among individuals with normoglycemia, prediabetes, and diabetes, respectively. In addition, ProAS were positively associated with CVD mortality and cancer mortality.

Fig. 2.

Associations of glycemic status-specific ProAS (A, normoglycemia; B, prediabetes; C, diabetes) (per SD) with mortality and incident diseases in the testing set. Models were adjusted for age, sex, race, Townsend deprivation index, BMI, never smoking, healthy drinking, physical activity, diet score, sleep score, cardiovascular disease, cancer, high cholesterol, hypertension, respiratory disease, antihypertensive, and cholesterol-lowering medications. In participants with diabetes, the model was additionally adjusted for diabetes duration, HbA1c, and diabetes medication. For incident diseases, only participants free from the corresponding disease at baseline were included and prevalence of corresponding disease was not included in the model. AF atrial fibrillation, BMI body mass index, CHD coronary heart disease, CKD chronic kidney disease, CVD cardiovascular disease, CI confidence interval, HR hazard ratio, HF heart failure, PAD peripheral artery disease, SD standard deviation

The Kaplan–Meier survival curves exhibited significant differences in mortality risk across the tertile groups of ProAS (Fig. 3A–C). The all-cause mortality rates of the highest tertile group were found to be significantly higher compared to those of the lowest tertile group, with an HR of 13.59 (95% CI, 10.58–17.47) in individuals with normoglycemia, 10.17 (6.72–15.39) in those with prediabetes, and 9.80 (6.06–15.83) in those with diabetes.

Fig. 3.

Predictive performance of different aging metrics for all-cause mortality stratified by glycemic status in the testing set. A–C Survival curves for all-cause mortality stratified by ProAS tertiles: top 25% (red), middle 50% (blue), and bottom 25% (green). Shaded regions represent standard errors derived from survival proportions. Hazard ratios (HR) for all-cause mortality with 95% Cis were estimated using Cox proportional hazards regression with the bottom 25% group as the reference. D–F ROC curves for 5-year, 10-year, and 15-year mortality risk prediction of four aging metrics. AUC area under the curve, FI frailty index, HR hazard ratio, TL telomere length, ROC receiver operating characteristic curve

Associations of glycemic status-specific ProAS with incident diseases

ProAS were significantly associated with all incident diseases included in this study (Fig. 2). Among participants with normoglycemia, each 1 SD increment in ProAS was associated with higher risks of CVD, CHD, stroke, HF, AF, CKD, PAD, cancer, and dementia after adjustment for potential confounders. The associations between ProAS and incident diseases were generally similar among participants with prediabetes and diabetes, while the most pronounced associations were found for PAD and CKD among participants with prediabetes and diabetes. The Kaplan–Meier survival curves showed significant differences in disease risk across the tertile groups of ProAS, with a progressive increase in HRs observed (Figure S5-Figure S7).

Comparisons of ProAS with other aging metrics

We compared ProAS with three well-known aging metrics (FI, TL, and chronological age) regarding their predictive ability for mortality risk and major disease incidence in the testing set. As shown in Fig. 3 and Figure S8, ProAS outperformed other aging metrics in predicting all-cause mortality during the entire follow-up period, irrespective of glycemic status (Fig. 3D–F and Figure S8). For 10-year all-cause mortality prediction, ProAS achieved an AUC of 0.777 in participants with normoglycemia, 0.767 in those with prediabetes, and 0.749 in those with diabetes, which was higher than chronological age, FI, and TL (P < 0.001).

Given the correlations of ProAS, FI, and TL with chronological age, we regressed each aging measure against chronological age and calculated the residuals to examine whether they retained predictive information independent of chronological age. It is evident that the residuals of the three biological aging metrics displayed a reduction in predictive ability of all-cause mortality during the follow-up period (Figure S9). However, the residuals of ProAS consistently outperformed those of other aging metrics across glycemic status. The age-stratified analyses further demonstrated that ProAS consistently outperformed other metrics in predicting all-cause mortality across age groups, underscoring its robustness beyond chronological age (Figure S10-S11).

We further examined the potential application of ProAS in predicting the risk of aging-related diseases and found that ProAS generally outperformed FI and TL, and either outperformed or showed similar accuracy to chronological age across glycemic status over 10-year incidence (Figure S12-Figure S14).

Discriminative improvement of ProAS beyond conventional predictors

Among participants with normoglycemia, ProAS alone displayed remarkable predictive ability for most endpoints. However, the predictive ability of ProAS decreased as glycemic status deteriorated, with an exception of PAD.

The differences in predictive performance by incorporating ProAS into the models with conventional risk factors are summarized in Fig. 4 and Table S10. Among participants with normoglycemia, ProAS added predictive utility of all-cause mortality and incident aging-related diseases, except for cancer, beyond conventional risk factors. Compared with the model that included conventional risk factors, the combined model integrating ProAS demonstrated a significant improvement in AUC of mortality and incident diseases. The better discriminative ability of ProAS was less evident in participants with impaired glucose status. The declined superior performance of ProAS across glycemic status was further confirmed by 10-year IDI and NDI (Table S11).

Fig. 4.

Predictive value offered by ProAS alone and in combination with other conventional risk factors for 10-year mortality and incident diseases stratified by glycemic status in the testing set. A–C AUCs of ProAS alone for predicting all-cause mortality and incident diseases. D–F Differences in AUC resulting from the addition of the ProAS to models with increasingly extensive sets of risk factors: age, sex, BMI, smoking status, alcohol drinking, physical activity, diet score, sleep score, SBP, total cholesterol, HbA1c, and creatinine. Individuals with a prevalent disease diagnosis before the baseline assessment were excluded from each disease-specific analysis. *Asterisk denotes a significant improvement in AUC with the addition of ProAS, as determined by DeLong’s test (p < 0.05). AUC the area under the curve, AF atrial fibrillation, BMI body mass index, CHD coronary heart disease, CKD chronic kidney disease, CVD cardiovascular disease, HF heart failure, HbA1c glycated hemoglobin, PAD peripheral artery disease, SBP systolic blood pressure

To test the predictive ability of individual proteins, we selected the top four important proteins identified within each glycemic status to further assess their prediction accuracy for all-cause mortality and incident diseases. As expected, GDF15 ranked highest in predicting all-cause mortality and most diseases among individuals with normoglycemia and prediabetes but its predictive ability diminished in those with diabetes, among whom WFDC2 was the leading predictor (Figure S15-S16). Adding these proteins into the conventional risk factors panel in individuals with corresponding glycemic status resulted in an AUC for all-cause mortality and incident diseases that were only slightly lower or comparable to that of combining ProAS with conventional risk factors (Figure S17).

Modifiable factors for ProAS

Among the 102 potentially modifiable factors analyzed from the baseline survey, 72, 51, and 36 factors showed a significant association with ProAS in participants with normoglycemia, prediabetes, and diabetes, respectively, after Bonferroni correction (Fig. 5 and Table S12). Both the number of significant modifiable factors and the strength of the observed associations decreased as glycemic status worsened. Notably, current smoking showed the strongest association with ProAS (vs never smoking, normoglycemia: β = 0.83, 95% CI: 0.80, 0.85; prediabetes: β = 0.46, 95% CI: 0.43, 0.49; diabetes: β = 0.41, 95% CI: 0.35, 0.46).

Fig. 5.

The modifiable factors for glycemic status-specific ProAS. Models were adjusted for age, sex, race, and antihypertensive and cholesterol-lowering medications. In participants with diabetes, the model was additionally adjusted for diabetes medication. The color of cells indicates the effect sizes (beta) between ProAS and each modifiable factor. Asterisks in cells represent significant associations after Bonferroni correction for multiple testing (p < 0.05/102 = 4.90×10 ). MET metabolic equivalent of task, SES socioeconomic status

Discussion

In this study, using a data-driven proteomic mapping approach, we delineated differential protein signatures that closely captured the risk of all-cause mortality among participants with varying glycemic status. Based on the distinct protein panels, we developed glycemic status-specific ProAS and found their robust associations with increased risks of mortality and aging-related diseases. Importantly, ProAS enhanced the predictive accuracy of multiple incident diseases beyond conventional risk factors but notable heterogeneity was observed across glycemic status. The number of diseases for which ProAS significantly yielded additional predictive accuracy decreased in individuals with prediabetes and diabetes as compared to those with normoglycemia. Next, we identified 72, 51, and 36 modifiable factors across seven categories that were significantly linked to ProAS in participants with normoglycemia, prediabetes, and diabetes, respectively, with current smoking being the most powerful factor for ProAS. Significantly, the strength of these associations progressively diminished as the glycemic status worsened.

Previous studies have suggested that impaired glycemic status, including prediabetes and diabetes, is associated with increased biological aging or brain aging [17, 18]. Compared to those without diabetes, individuals with prediabetes and diabetes have been shown to have an average increase in biological age calculated from clinical biomarkers of 2.69 and 12.02 years, respectively [17]. Similarly, brain age assessed by MRI-based brain volume and functional connectivity measures was on average 0.50 years older than chronological age among participants with prediabetes and 2.29 years older among participants with diabetes [18]. These findings indicate the heterogeneity of biological aging and susceptibility to aging-related diseases among individuals with different glycemic status. It is noteworthy that although HbA1c or glucose has been previously incorporated into biological age estimation [36, 37], the role of protein metabolism in accelerated aging process associated with diabetes or prediabetes is largely overlooked [38, 39]. Proteins, as the final products of gene expression, may provide a more direct mechanistic and functional insight into aging biology. Several previous studies have successfully developed proteomic aging clocks that exhibited significant association with mortality and common diseases [11, 14]. However, their applicability to glycemic status-specific aging patterns has not yet been evaluated, thereby impeding their capacities to precisely capture aging patterns and related outcomes for vulnerable population, such as patients with diabetes. In this case, quantifying proteomic profiles according to glycemic status would contribute to tailored prevention and intervention strategies to promote healthy aging.

The 577, 302, and 333 proteomic aging biomarkers identified among individuals with normoglycemia, prediabetes, and diabetes in our study were found to be involved in a wide range of biological processes, primarily including immune function, cell proliferation, differentiation and survival, as well as inflammation, suggesting that aging is a complex, multidimensional phenomenon. Using the LASSO and LightGBM model, we ultimately identified a panel of 11, 23, and 21 important proteins among participants with normoglycemia, prediabetes, and diabetes, respectively. Among these, GDF15 was the top-ranked predictor of all-cause mortality in normoglycemia and prediabetes, but its predictive ability diminished in diabetes, where WFDC2 emerged as the leading predictor. GDF15, a stress-responsive mitokine, protects against aging-mediated inflammation and regulates endothelial function and insulin sensitivity via activation of the PI3K/Akt pathway and inhibition of NF-κB/JNK signaling [40]. WFDC2, also known as human epididymal protein 4 (HE4), is a secreted glycoprotein implicated in profibrotic processes, particularly renal fibrosis, by inhibiting multiple proteases and limiting type I collagen degradation, which promotes extracellular matrix accumulation and accelerates renal decline in diabetes [41, 42]. Although they function via largely independent mechanisms, GDF15 and WFDC2 may converge indirectly in inflammation- and apoptosis-related pathways as nonspecific markers of multi-organ tissue damage [43, 44], upregulated in response to stressors such as inflammation, hypoxia, and oxidative stress, thereby contributing to aging and chronic disease progression [45]. It is worth noting that WFDC2 became increasingly important for mortality prediction with the progression of glucose regulation disorders. This situation is similar for several other proteins such as NT-proBNP and CLEC3B. NT-ProBNP is a cardiac hormone secreted by cardiomyocytes in response to increased ventricular wall stretch [46], and CLEC3B/tetranectin as a regulator of the fibrinolysis and proteolytic system has been shown to exacerbate type 2 diabetes by inhibiting insulin secretion from β cells [47]. The observed differences in protein signatures across glycemic states likely reflect the progressive impact of hyperglycemia on organ-specific vulnerabilities and systemic dysfunction [48, 49]. In normoglycemia and prediabetes, proteins like GDF15 may play an important role in mitigating aging‐related inflammation to maintain glucose homeostasis and insulin sensitivity [50]. However, the dominance of WFDC2, NT-proBNP, and CLEC3B in diabetes reflects the specific pathologies under disturbed glucose metabolism. WFDC2 is linked to renal fibrosis that underscores accelerated kidney damage from chronic hyperglycemia [41]. Elevated NT-proBNP signals heightened ventricular stress and cardiac strain associated with diabetic cardiomyopathy or vascular complications [46, 51]. The involvement of CLEC3B in β-cell dysfunction points to worsening insulin resistance and pancreatic exhaustion [47]. These shifts suggest that as glycemic status deteriorates, proteomic aging patterns capture more about end-organ damage and metabolic decompensation rather than general inflammatory processes, underscoring the need for developing aging clocks stratified by glycemic status [52].

We then developed glycemic status-specific ProAS by integrating the identified protein biomarkers and found that ProAS was strongly associated with the risks of mortality and various incident diseases. Moreover, ProAS showed superior performance in predicting all-cause mortality and diseases over several well-established aging measures across different follow-up periods. These findings underscore the considerable potential of ProAS in early identification of individuals at high risk and enabling timely interventions. Notably, our study adds to the existing evidence on the utility of proteomic aging clock by demonstrating that the predictive performance of ProAS declined for incident diseases as glycemic status deteriorated, with an exception of PAD. In addition, we detected a significant improvement in the predictive capacity for all-cause mortality and multiple diseases beyond conventional risk factors. However, the number of diseases for which ProAS exhibited an additional predictive utility diminished among participants with prediabetes and diabetes compared to those with normoglycemia. The lower predictive utility of ProAS among individuals with abnormal glucose regulation might be partly attributed to the complexity of physiological responses to nutritional, exercise, and pharmacological interventions. Collectively, our findings imply that glycemic status should be taken into consideration when evaluating the potential benefits of adopting proteomic aging signature as a valuable tool to empower comprehensive risk assessments for death and diseases.

Interventions targeting the aging process have the potential to prevent and delay the onset of age-related diseases [53]. Proteomic aging measures could serve as a surrogate for assessing the effectiveness of interventions. Recently, Wang et al. reported associations of smoking, BMI, diabetes, hypertension, and CVD with proteomic aging in the Atherosclerosis Risk in Communities study [15]. In our study, we conducted a comprehensive investigation of the modifiable factors for glycemic status-specific proteomic aging, and eventually identified 72, 51, and 36 potential factors in participants with normoglycemia, prediabetes, and diabetes, respectively, across seven categories including local environment, psychosocial factors, socioeconomic status, medical history, early life factors, physical measures, and lifestyle. Among them, current smoking seemed to be the most influential factor. Our study provides a proteomic insight into the mechanisms linking these factors to aging process, underscoring the significance of targeted interventions to slow down the aging process and reduce the risks of mortality and morbidity in later life. Furthermore, both the number and magnitudes of the observed associations between modifiable factors and proteomic aging decreased with worsening glycemic status. The attenuated impact of potential factors on proteomic aging in individuals with prediabetes and diabetes may be related to the cumulative effect of hyperglycemia and concomitant metabolic changes. Chronic hyperglycemia induces oxidative stress, systemic inflammation, and the accumulation of advanced glycation end-products, which in turn accelerate cellular and metabolic dysfunction [54–56] and potentially overshadow the effects of modifiable factors on biological aging.

To our knowledge, this is the first study to systematically characterize proteomics-based aging patterns across varying glycemic statuses and their predictive value for health outcomes using a holistic proteomics strategy, which contributes to refining risk stratification and implementing targeted interventions. In people with prediabetes, ProAS may help identify individuals with accelerated aging who are likely to benefit from risk factor control for the prevention of diabetes or other age-related diseases. While in people with diabetes, ProAS may help inform tailored interventions for individuals who are at risk of complications and assess the efficacy of diabetes management. Targeting lifestyle modification, including smoking cessation, might be an effective and feasible strategy to help achieve healthy aging.

The strengths of our study include the large-scale prospective cohort, the long-term follow-up, and high-throughput proteomics measurements. However, several limitations within this study should be noted. First, the plasma proteome currently available in the UK Biobank is based on the Olink Explore 3072 assay and, therefore, failed to capture all proteins covered in other platforms and panels (e.g., SOMAscan). Second, due to the lack of data on fasting glucose and postload glucose, screening-detected prediabetes and diabetes were determined based on HbA_1c measurements only. Therefore, a proportion of participants with prediabetes or diabetes might have been misclassified as normal or diabetes-free, possibly leading to a dilution of the observed differences. HbA1c has several well-recognized limitations: it can be affected by red blood cell turnover, certain comorbidities (e.g., advanced kidney or liver disease), genetic hemoglobin variants, and acute or exogenous factors such as medications or transient illness [57, 58]. Moreover, discrepancies between HbA1c and other glycemic measures, such as fasting glucose or oral glucose tolerance tests, may further contribute to misclassification, particularly among individuals near diagnostic thresholds [59]. These considerations underscore the need for cautious interpretation of our findings and, where feasible, the integration of multi-modal glycemic assessment in future studies. Third, the associations between modifiable risk factors and ProAS are cross-sectional, and causal inferences could not be made accordingly. Moreover, although we included numerous modifiable factors for analysis, it is conceivable that certain factors may have been unintentionally omitted. Fourth, the absence of proteomic profiling at multiple time points precludes us from calculating proteomic aging rates, which may offer complementary information to unveil the dynamic progression of aging process. Finally, the UK Biobank cohort is predominantly composed of White British and subject to healthy volunteer bias. Therefore, our results should be generalized and interpreted with caution.

Conclusions

In conclusion, our study initially identified glycemic status-specific protein signatures associated with mortality through large-scale proteomic mapping. The ProAS, built on corresponding protein signatures, demonstrated strong predictive power of mortality and multiple diseases up to 15 years of follow-up. Of particular interest is the fact that the predictive performance of ProAS for a majority of diseases declined across worsening glycemic status. Glycemic status-related differences in modifiable factors for proteomic aging were also observed. Our findings extend existing aging clocks by evaluating proteomic aging patterns and their capacity to predict health outcomes across distinct glycemic subgroups, providing critical insights for the development of tailored prevention and intervention strategies, although external validations in independent cohorts are further needed.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Bin Wang and Yingli Lu conceived and designed the study. Jiang Li and Jie Li performed the statistical analysis and drafted the manuscript. Xiaoqin Xu, Yuefeng Yu, Wenqi Shen, Ying Sun, and Yanqi Fu participated in data collection. Xiao Tan and Ningjian Wang critically revised the manuscript. All authors read and approved the final manuscript. Bin Wang is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Funding

This study was supported by the National Natural Science Foundation of China (82170870 and 82120108008), the Science and Technology Commission of Shanghai Municipality (22015810500), and the Major Science and Technology Innovation Program of Shanghai Municipal Education Commission (2019–01-07–00-01-E00059). The funders of the study had no role in study design, data collection, analysis, decision to publish, or manuscript preparation.

Data Availability

For the UK Biobank study, the datasets generated and analyzed are available at https://www.ukbiobank.ac.uk/, and our research has been conducted under application number 77740.

Declarations

Competing interests

The authors declare no competing interests.