Proteomics-based aging clocks in midlife or late-life and their associated risk of dementia

Abstract

Background:

Biological age can be quantified by composite proteomic scores, called proteomics-based aging clocks (PACs). We investigated whether a discrepancy between chronological and biological age in midlife and late-life is associated with cognition and dementia risk.

Methods:

We used two longitudinal population-based studies: the Atherosclerosis Risk in Communities (ARIC) Study and the Multi-Ethnic Study of Atherosclerosis (MESA). PACs were created in ARIC at midlife (mean age: 58 years, 57% female, n = 11,758) and late-life (mean age: 77 years, 56% female, n = 4934) using elastic net regression models in two-thirds of dementia-free participants and validated in the remaining one-third of participants. Proteomics-based age acceleration (PAA) was calculated as residuals after regressing PACs on chronological age. We validated the midlife PAC in the MESA cohort (mean age: 62 years, 52% female, n = 5829). We used multivariable linear and Cox proportional hazards regression to assess the association of PAA with cognitive function and dementia incidence, respectively.

Results:

In ARIC, every five years, PAA is associated with lower global cognition: difference: −0.11, 95% confidence interval[CI]: −0.16, −0.06) using midlife PAA and difference: −0.17, CI: −0.23, −0.12 using late-life PAA. Midlife PAA is associated with higher dementia risk (hazard ratio[HR]: 1.20 [CI: 1.04, 1.36]) and more prominently when using late-life PAA (HR: 2.14 [CI:1.67, 2.73]). Similar findings are observed in MESA: PAA is associated with lower global cognitive function (difference: −0.08 [CI: −0.14, −0.03]) and higher dementia risk (HR:1.23 [CI: 1.04, 1.46]).

Conclusions

Accelerated biological age is associated with lower cognition and a higher risk of dementia in midlife and more prominently in late life.

Plain language summary

We studied whether people who age faster biologically than their calendar age have worse cognitive function and a higher risk of dementia. Using two large population-based studies, we created a score based on blood proteins to measure biological age in midlife and later life. We found that when individuals’ biological age was higher than their calendar age, they tended to have poorer cognitive function and were more likely to develop dementia. In the future, this finding could help identify people who might benefit from early lifestyle or medical interventions to keep their brains healthy.

Sedaghat et al. examine the association between proteomics-based aging clocks (PAC) during midlife and late-life and the risk of dementia in individuals from two large prospective population-based cohort studies. They find that the PACs are associated with lower cognition and a higher risk of dementia in midlife and more prominently in late-life.

Introduction

Dementia is a major cause of death, disability, and dependency among older adults worldwide^1,2. While age is the main risk factor for dementia, it is well recognized that biological aging differs between individuals³. Biological age can deviate from chronological age due to various biological disruptions such as inflammation, oxidative stress, vascular dysfunction, and immune dysregulation³. These biological disturbances typically become more prominent with advancing age, but they vary among individuals and are referred to as accelerated biological aging⁴.

Several studies have shown that patients with dementia develop subclinical metabolic changes years before dementia onset^5,6. Studies have demonstrated that there are significant differences in plasma biomarker composition between patients with dementia and cognitively intact individuals^5,7. This finding has stimulated various lines of research to build up biological aging clocks that can predict future decline in brain structural and functional integrity⁸. Specifically, biological aging processes can be quantified using composite metrics referred to as aging clocks using plasma protein biomarkers³. Proteomics-based aging clocks (PACs) are promising biomarkers of aging because protein expression changes with advancing age and they exert biological functions that can be potentially modified by lifestyle and pharmaceutical interventions^3,9. In this study, we created PACs across two life stages, namely, midlife and late-life, then tested the hypotheses that PACs at each stage would be associated with lower global and domain-specific cognitive function and increased dementia risk. We hypothesized that late-life PAC would show a stronger association as it may capture greater variability in protein levels in older age. We performed this study in the setting of two large prospective population-based cohort studies. PACs were created and tested in the Atherosclerosis Risk in the Communities (ARIC) study, a cohort of mostly White and Black men and women with plasma proteomics data that have been collected over 20 years of follow-up, and then validated and replicated in the Multi-Ethnic Study of Atherosclerosis (MESA), a prospective cohort including ethnically diverse participants identifying as Black, White, Asian, and Hispanic. We find that higher PAC is associated with lower cognitive function and higher risk of dementia, and the associations are stronger when using late-life PAC.

Methods

The ARIC study population

The ARIC study is a longstanding prospective cohort of 15,792 participants (45–64 years old) started in 1987–1989 (Visit 1)^10–12. Participants were recruited from four communities in the United States (suburban Minneapolis, MN; Washington County, MD; Forsyth County, NC; and Jackson, MS)¹³. Participants have been re-invited for follow-up visits, including Visit 2 (1990–1992), Visit 3 (1993–1995), Visit 4 (1996–1998), Visit 5 (2011–2013), Visit 6 (2016–2017), and Visit 7 (2018–2019), of relevance to this analysis. We included 11,758 participants (57% female) with information on cognitive function and protein measurements at Visit 2 (midlife) and 4,934 participants (56% female) at Visit 5 (late life) (Fig. S1a, b) to train midlife and late-life PACs and analyze their association with incident dementia.

We also investigated the association between PACs and cognitive function. We included 5,123 participants at Visit 5 (late-life) who had available information on both cognitive function and proteins (Fig. S2). There were 4,783 participants with protein data at Visit 2 (midlife) and cognitive function data at Visit 5 (Fig. S2).

The study was approved by each site’s institutional review board, and written informed consent was signed by all participants (or proxies, when required). For this study, we obtained anonymized data through manuscript proposal submissions to the publication committees of both cohorts. Additional ethical review was not required according to the policies of our institute.

Proteomics measurement

In ARIC, plasma proteins have been measured using a SOMAmer (Slow Off-rate Modified Aptamers)-based assay called SomaScan (V4.0) (SomaLogic, Inc., USA)¹⁴ in stored blood samples collected at Visit 2 (midlife) and Visit 5 (late-life). The SomaScan platform uses single-stranded DNA-based protein-bound aptamers to capture conformational protein epitopes^14,15. The aptamers are mapped to unique proteins using the Universal Protein Resource (UniProt) databases^15,16. Approximately 5000 proteins (4955 aptamers and 4712 unique proteins) measured at midlife and late-life underwent SomaScan standardization and normalization processing^11,17. Briefly, hybridization control normalization was applied to each sample to correct systematic biases, followed by median signal normalization to eliminate sample or assay biases within plates. Based on global reference plate-to-plate variations were assessed and protein analytes with a calibration factor ±0.4 (the median calibration factor) were excluded from all analyses. This process was used to ensure minimal batch effect and absence of systematic biases when using proteins from different visits longitudinally. To correct for skewness, all aptamer measures were log base 2 transformed. We ran blind split-sample duplicate plasma aliquots and observed median coefficient variation of 6% and 7%, and median Pearson correlations of 0.93 and 0.96 at midlife (Visit 2) and late-life (Visit 5), respectively.

Cognitive function assessment

All participants completed a 60-min comprehensive neuropsychological assessment administered by trained and certified psychometrists at Visit 5 (late-life). The measures are well-validated and standardized instruments, which assess multiple domains of cognition, including memory, executive function/processing speed, and global cognitive function¹⁸. The test battery includes: Memory domain: Delayed Word Recall Test: a test of verbal memory requiring recall of a word list after a short delay (score range 0–10). Logical Memory I and II: from the Wechsler Memory Scale-Revised (WMS-R) is a test of immediate (Logical Memory I) and delayed (Logical Memory II) memory. Executive function/processing speed domain: Trail Making Test Part A: In Trail Making A, participants are asked to draw a line connecting circles numbered 1 to 25 that are randomly distributed on the page as fast as possible. Digit Span Backwards: a test of attention in which participants state a series of digits backward. Digit Symbol Substitution Test: a subtest of the Wechsler Adult Intelligence Scale-Revised involving timed translation of numbers to symbols in 90 seconds using a key, which measures psychomotor performance (score range 0–93). Word Fluency Test: combined total of correct words produced beginning with F, A, and S. Category Fluency Test: participant is asked to spontaneously generate words from a specific category (in this test, animals)^18,19. For global cognitive function assessment, we included all the aforementioned tests as well as the Boston Naming Test: a test of language in which participants name common objects from pictures. To create scores for each cognitive domain and global cognitive function, we used principal component analysis (PCA) to derive three cognitive function scores for memory, executive function/processing speed, and global cognition (combination of all cognitive domains). Before PCA analysis, cognitive function test scores were checked for normal distribution. Trails A test scores were inverted so that low test scores indicate poorer cognitive function and higher test scores indicate better cognitive function for all tests. Participants with no cognitive function scores were excluded (N = 60). Imputation via mean was used to impute any missing values for cognitive tests (total number of missing in any test = 439). Next, Z-scores were calculated for all cognitive test scores, and PCA was conducted to create three distinct factor scores for memory, executive function/processing speed, and global cognition. Percentages of variance explained by PCA factors for each cognitive domain are compiled in Table S1.

Dementia incidence

Dementia incidence was assessed using well-validated, standardized battery of cognitive measures supplemented by dementia surveillance in between visits, and hospital discharge or death certificate^18,20–23. In short, all participants underwent a 3-instrument cognitive assessment at Visit 2, Visit 3, and Visit 4. The 3-instrument cognitive testing was repeated in a subset at the ARIC-MRI examination in 2004–2006 (Jackson and Forsyth County sites only); and again, in all participants who took part in in-person assessments at Visits 5, 6 and 7 as part of the ARIC-NCS (NeuroCognitive Study). From Visit 5 onwards, those unwilling or unable to attend the in-clinic assessment were invited for an in-person assessment in their home or long-term care facility. If they did not take place in-person in visit 5, they were offered a modified telephone interview for cognitive status (TICSm). Beginning in 2012, participants were screened for dementia on annual or semi-annual cohort follow-up calls using the Six-Item Screener, then for those with an indication of impaired cognition, the Eight-item Informant Interview to Differentiate Aging and Dementia (AD8) was conducted with proxies. The data was supplemented by International Statistical Classification of Diseases and Related Health Problems (ICD) codes for dementia identified through surveillance of hospital discharges or death certificates. The information on dementia was reviewed according to a standard protocol by the ARIC Neurocognitive Classification Committee. The dementia onset was the earliest date determined by in-person visit assessment, dementia surveillance, hospital discharge, or death certificate code. When dementia was identified through an informant interview, hospitalization record, or death certificate, the date of diagnosis was estimated to occur 180 days before the documented incident or interview¹⁸. Follow-up time was defined as the number of days from the participant’s baseline exam to the date of incident dementia event, loss to follow-up, death from another cause, or censoring date at December 31st, 2019, whichever occurred first.

Other covariates

All covariates were assessed at visits where proteins were measured (ARIC: Visit 2, midlife, and Visit 5, late-life). Race was self-reported and was classified as Black or White. Cigarette smoking and education were assessed using questionnaires and were categorized as current, former, or never users for smoking status and less than completed high school, high school equivalent, and greater than high school for education (measured at Visit 1). Diabetes was defined as self-reported history of physician diagnosis, antidiabetic medication use during the past 2 weeks, fasting blood glucose level ≥ 126 mg/dL, or nonfasting blood glucose level ≥ 200 mg/dL. Trained technicians measured blood pressure with participants sitting after 5-min rest. Blood pressure was measured three times using and the average of the last two readings was recorded. Hypertension was defined as systolic blood pressure greater than 140 mm Hg or diastolic blood pressure greater than 90 mm Hg or using antihypertensive medications. Plasma total cholesterol and creatinine, and cystatin C were measured using enzymatic methods. Estimated glomerular filtration rate (eGFR) was calculated using CKD-EPI 2021 equation²⁴. Information on physical activity was collected using the interviewer-administered Baecke questionnaire. Total metabolic equivalent minutes per week at Visit 1 and 5 were used as a continuous covariate in the midlife and late-life models, respectively^25,26. Information regarding diet was collected using the interviewer-administered food-frequency questionnaire at Visits 1 only. We used the American Heart Association healthy diet score, incorporating intake of fruits and vegetables, fish, whole grains, and sugar-sweetened beverages, while sodium intake was excluded due to data unavailability^27,28. This score was modeled as a categorical variable for both midlife and late life, with three levels of diet quality: poor, intermediate, and ideal. Social isolation was assessed at Visit 2 using the Lubben Social Network Scale (LSNS), which asks about the size of the participant’s active social network and the perceived availability of social support. The total score from the LSNS (range: 0–50) was modeled as a categorical variable for both midlife and late-life with four groups based on the Lubben criteria, a commonly used method in ARIC studies: ≤20 = isolated; 21–25 = high risk for isolation; 26–30 = moderate risk for isolation; ≥31 = low risk for isolation^29,30. Traumatic brain injury (TBI) at Visit 2 and Visit 5 was defined based on self-report and data from emergency department visits and inpatient hospitalizations using ICD-9 or ICD-10 codes^31,32. Genotyping for APOE was performed by TaqMan assay (Applied Biosystems, Foster City, Calif).

Validation of PACs in the MESA cohort

The MESA cohort included 6,814 men and women aged between 45 and 84 who identified their race/ethnicity as White, Black, Chinese, or Hispanic/Latino and who had no history of clinical cardiovascular disease at enrollment. Participants were recruited from Baltimore, Maryland; Chicago, Illinois; Forsyth County, North Carolina; Los Angeles County, California; New York, New York; and St. Paul, Minnesota. Enrollment and the baseline exam (Exam 1) occurred between 2000 and 2002. Participants have been re-invited for follow-up visits, including Exam 2 (2002–2004), Exam 3 (2004–2005), Exam 4 (2005–2007), Exam 5 (2010–2011), Exam 6 (2016–2018), and Exam 7 (2022–2024)³³. We included 4057 participants (52% female) with protein measurements at Exam 1 and cognitive function information at Exam 5 and 5829 participants with both protein measurements and information on dementia status (Fig. S3). The institutional review boards at all participating sites approved the study, and all participants provided written informed consent.

Proteins have been measured using a newer SomaScan version (V4.1), including 7000 proteins. This version contains all 5000 proteins from the previous version of the SomaScan assay that was used in ARIC. We used the same quality checks and protocols as in ARIC. To confirm similar measurements of proteins in MESA and ARIC cohorts, we compared the distribution and summary statistics of aptamers between the studies; 4 random aptamers (protein units) are presented in Fig. S4. Cognitive function was measured in a research setting at Exam 5 (2010–2011). The battery includes the Cognitive Abilities Screening Instrument (CASI), digit symbol coding, forward digit span, and backwards digit span. General instructions for the cognitive examination were translated into Spanish and Mandarin Chinese and then independently back-translated by native speakers and pretested³⁴. We used individual cognitive tests in MESA. Incident dementia was ascertained through ICD-9 or ICD-10 codes in medical records for hospitalizations reported during follow-up interviews, as well as in dementia death certificates. The codes used to define dementia have been listed previously³⁵. Follow-up time was defined as the number of days from the participant’s baseline exam to the date of incident dementia event, loss to follow-up, death from another cause, or censoring date at December 31st, 2018, whichever occurred first.

Statistics and reproducibility

Proteomic aging clocks (PACs) and proteomics-based age accelerations (PAA)

We created and trained the PACs in the ARIC cohort. To construct midlife and late-life ARIC dementia-free PACs, we randomly selected two-thirds of participants who remained free of dementia until 2019 at each midlife and late-life visit and used them as the training set at the corresponding visits. The remaining one-third of participants who remained free of dementia until 2019 were used as the test set at the corresponding visits (Fig. S1a, b). Using the training set, we applied elastic net regression to train the ARIC dementia-free PACs against age as a weighted sum of aptamers: $chronologicalage={\beta }_0+{\sum }_{i=1}^n{\beta }_i{\times aptamer}_i$, where aptamer_i is the level of the ith aptamer³⁶. Lambda value was selected based on 10-fold cross-validation in the training set. This resulted in the selection of 1176 aptamers in ARIC midlife and 618 aptamers in ARIC late-life participants (Supplementary Data 1 and 2). We internally validated the ARIC dementia-free PACs by examining their correlation with age in the remaining participants at the corresponding visits. To capture PACs’ effect independent of age, we created proteomics-based biological age acceleration (PAA) for each PAC as residuals by regressing PAC on chronological age in the remaining participants after excluding the training set at the corresponding visits. A positive value of PAA suggests that the proteomic age tends to be older than the person’s chronological age (Fig. 1). To understand the combination of proteins contributing to midlife and late-life PACs, we took a closer look at the overlapping aptamers (Fig. S5, Supplementary Data 3). There were 270 overlapping aptamers between midlife and late-life. At both timepoints, we selected the top 10 proteins based on their effect estimates with chronological age (derived from the elastic net regression) for presentation herein. Among them, 6 were both at midlife and late-life, so a total of 14 proteins are presented Table 1).

Fig. 1. Proteomics-based aging clock training and validation.

Using a random 2/3 of dementia-free participants in ARIC, proteomics-based aging clocks were trained using elastic net regression against chronological age. PACs were created as a weighted sum of aptamers (protein units): $chronologicalage={\beta }_0+{\sum }_{i=1}^n{\beta }_i{\times aptamer}_i$. In the remaining 1/3 of participants and those with dementia, proteomics-based aging clocks were created using weights derived from the training set and individuals’ protein values. Weights are regression coefficients from the elastic net regression models against chronological age. PAC proteomics-based aging clock, MESA Multi-Ethnic Study of Atherosclerosis, ARIC Atherosclerosis Risk in Communities. Created in BioRender. (2025) https://BioRender.com/llf8pf4.

Table 1.

Top (based on effect estimates with chronological age) overlapping^a proteins between midlife and late-life

Protein Name	Mechanisms of action	Midlife effect estimatesb	Late-life effect estimatesb
Prostaglandin-H2 D-isomerase (HPGDS)	Catalyzes both the conversion of Prostaglandin-H2 to Prostaglandin D2, a prostaglandin involved in smooth muscle contraction/relaxation and a potent inhibitor of platelet aggregation, and the conjugation of glutathione with a wide range of aryl halides and organic isothiocyanates. Also exhibits low glutathione-peroxidase activity towards cumene hydroperoxide.	1.09	0.36
Lumican (LUM)	Belongs to the small leucine-rich proteoglycan (SLRP) family.	1.48	0.31
Transgelin (TAGLN)	Involved in calcium interactions and contractile properties of the cell that may contribute to replicative senescence; Belongs to the calponin family.	1.60	1.38
CUB domain-containing protein 1 (CDCP1)	Belongs to the tetraspanin web involved in tumor progression and metastasis.	0.75	0.51
Pleiotrophin (PTN)	Regulates many processes like cell proliferation, cell survival, cell growth, cell differentiation, and cell migration in several tissues, namely neurons and bone.	1.68	2.16
Coiled-coil domain-containing protein 80 (CCDC80)	Promotes cell adhesion and matrix assembly.	1.00	0.43
Chordin-like protein 1 (CHRDL1)	Antagonizes the function of Bone Morphogenetic Protein 4 by binding to it and preventing its interaction with receptors. Alters the fate commitment of neural stem cells from gliogenesis to neurogenesis. Contributes to neuronal differentiation of neural stem cells in the brain by preventing the adoption of a glial fate.	1.95	2.25
Growth/differentiation factor 15 (GDF15)	Regulates food intake, energy expenditure, and body weight in response to metabolic and toxin-induced stresses.	1.01	0.06
Neurogenic locus notch homolog protein 3 (NOTCH3)	Functions as a receptor for membrane-bound ligands Jagged1, Jagged2 and Delta1 to regulate cell-fate determination. Upon ligand activation through the released notch intracellular domain (NICD). It forms a transcriptional activator complex with RBPJ/RBPSUH (Recombination Signal Binding Protein for Immunoglobulin Kappa J region) and activates genes of the enhancer of split locus. Affects the implementation of differentiation, proliferation and apoptotic programs (By similarity).	0.93	1.33
Cartilage acidic protein 1 (CRTAC1)	Inhibits cell proliferation, migration, invasion and epithelial-mesenchymal transition (EMT) process in bladder cancer by downregulating Yin Yang 1 (YY1) to inactivate the transforming growth factor beta pathway.	0.97	0.76
WNT1-inducible-signaling pathway protein 2 (CCN5)	Promotes the adhesion of osteoblast cells and inhibits the binding of fibrinogen to integrin receptors.	1.58	1.83
R-spondin-4 (RSPO4)	Activator of the canonical Wnt signaling pathway by acting as a ligand for Leucine-rich repeat-containing G-protein-coupled receptors 4–6 (LGR4-6) receptors. Also regulates the canonical Wnt/beta-catenin-dependent pathway and non-canonical Wnt signaling by acting as an inhibitor of Zinc And Ring Finger 3 (ZNRF3).	0.59	0.58
EGF-containing fibulin-like extracellular matrix protein 1 (EFEMP1)	Binds epidermal growth factor receptor (EGFR), the receptor for epidermal growth factor (EGF), inducing EGFR autophosphorylation and the activation of downstream signaling pathways. May play a role in cell adhesion and migration.	0.51	0.62
Scavenger receptor class F member 2 (SCARF2)	Mediates homophilic and heterophilic interactions.	2.37	1.21

^aFrom 270 overlapping proteins between midlife and late-life, top 10 proteins based on effect estimates were selected at both midlife and late-life. Among them 6 (Pleiotrophin, Chordin-like protein 1, Cartilage acidic protein 1, WNT1-inducible-signaling pathway protein 2, Scavenger receptor class F member 2, and Transgelin) were both at midlife and late-life, which ended in 14 proteins presented here. S 3 includes full list of overlapping proteins.

^bEffect estimates are from the associations with age at each life stage from elastic net regression models using ~5000 proteins as predictors at each life stage and chronological age as the outcome in the training set.

PAA and cognitive function

Multivariate linear regression was used to calculate adjusted effect estimates and 95% confidence intervals (CI) for PAA (per 5 years) and global and cognitive function domain scores. The cross-sectional analyses were run using ARIC Visit 2 (midlife) and Visit 5 (late-life) PAA measures and cognitive function assessed at Visit 5 (late-life) (Fig. S6a). For all analyses, we ran two models: the first model adjusted for chronological age, sex, race/ethnicity, study center, and the second model additionally adjusted for education, body mass index (BMI), smoking status, hypertension, diabetes status, cholesterol, and eGFR. To contextualize the effect estimates of PAA on cognitive function, we conducted parallel analyses using APOE genotype as the exposure. Specifically, we compared individuals with one or two ε4 alleles to those with the ε3/ε3 genotype in relation to global cognitive function.

PAA and dementia

We used Cox proportional hazards regression models to examine the longitudinal association of PAA (per 5 years) with incident dementia. Analyses include participants from ARIC Visit 2 (midlife) and ARIC Visit 5 (late-life) to the date of incident dementia event, loss to follow-up, death from another cause, or end of follow-up (Fig. S6a). We ran the analyses in the remaining participants after excluding the training set at each visit. Because the remaining set included those who had dementia in ARIC, we applied a case-cohort weighting scheme employing Barlow’s method to account for the imbalance of dementia-free (one-third) and dementia (100%) participants³⁷. Following the case-cohort analysis method, we created a “subcohort” which consisted of the participants who were free of dementia in the remaining set and one-third randomly selected participants who developed dementia during follow-up, since the training and test sets split was 2:1. We ran all analyses in two adjustment models as mentioned before. The proportional hazards assumptions were not violated, as assessed by visual inspection of the survival curves and by assessing Schoenfeld residuals.

External validation in MESA

PACs were computed by multiplying the concentration of log 2-transformed proteins at Exam 1 by regression coefficients (weights) calculated in ARIC. The distribution of age at ARIC midlife and MESA Exam 1 are similar, while the late-life ARIC population is on average older than the MESA Exam 1 participants; therefore, we used PAC created at ARIC Visit 2 (midlife) as our primary clock for replication in MESA and PAC at ARIC Visit 5 (late-life) as a secondary clock (Fig. 1). The performance of the PACs were tested by (1) plotting the chronological age against PACs and (2) calculating median absolute error and Pearson correlation (r) with chronological age between PAC and chronological age. Ideally, r should be > 0.7 (Table S2). We then calculated PAA as residuals of PAC regressed on chronological age.

Multivariate linear regression was used to calculate adjusted effect estimates and 95% confidence intervals (CI) for the cross-sectional association of the PAA and individual cognitive function scores. We used PAA calculated at MESA Exam 1 (2000–2002) and cognitive function at Exam 5 (2010–2011). We used Cox proportional hazard regression models to examine the longitudinal association of PAA (per 5 years) at Exam 1 with the incidence of dementia (Fig. S6b). All analyses were done in two models adjusted for similar covariates as ARIC cohort. We repeated the analyses in MESA using proteins and coefficients based on ARIC late-life PACs (instead of ARIC midlife) in association with cognitive function and dementia incidence.

Sensitivity analyses in the ARIC cohort

To ensure cognitive imputation did not affect the results, we repeated the analyses of PAA and cognitive function using non-imputed data. To better understand the differences between midlife and late-life PACs, we re-created a PAC in ARIC late-life by using protein selection and regression coefficients (weights) from ARIC midlife and applying them to protein levels from late-life. In another sensitivity analyses, we additionally adjusted the analysis of dementia incidence for additional dementia modifiable risk factors, including depression, physical activity, American Heart Association diet score, social isolation, and traumatic brain injury. To explore whether the results are different based on race, APOE ε4 carriership (carrying 1 or 2 ε4 alleles), and sex, we stratified based on race, APOE ε4 allele carriership (carrying 1 or 2 ε4 alleles compared with no ε4 allele), and sex.

Reporting summary

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

Results

Baseline characteristics of participants in midlife and late-life cohorts in ARIC as well as MESA are presented in Table 2. Participant’s characteristics in those with and without incident dementia are presented in Tables S3 and 4. In addition, baseline characteristics of participants in the subset of those with cognitive function data are presented in Table S5. Compared to those included, individuals excluded from the analysis at Visit 2 were slightly younger, but there was no significant difference in cardiovascular risk profile, whereas those excluded from the analysis at Visit 5 were older and had a less favorable cardiovascular risk profile (Tables S6 and 7). PAA ranged in ARIC from −11.5 to 16.8 years at midlife and −7.4 to 12.5 years at late-life. In MESA, PAA ranged from −12.4 to 17.0 years. For dementia incidence, median follow-up time since ARIC Visit 2 (midlife) and Visit 5 (late life) were 21 years (interquartile range: 11) and 6 years (interquartile range: 3), respectively. In MESA median follow-up time was 17 years (Interquartile range: 5).

Table 2.

Characteristics at midlife (1990–1992) and late-life (2011–2013) in ARIC and Exam 1 (2000–2002) in MESA

	ARIC		MESA
	Midlife	Late-life
Characteristics	Total (N = 5420)	Total (N = 2116)	Total (N = 5829)
Demographics
Mean age, years (SD)	58.3 (5.7)	76.8 (5.3)	62.1 (10.3)
Female (N, %)	3088 (57.0)	1196 (56.5)	3033 (52.0)
Race (N, %)
Black	1322 (24.4)	420 (19.8)	1525 (26.1)
White	4098 (75.6)	1696 (80.2)	2301 (39.5)
Hispanic/Latino	–	–	1305 (22.4)
Chinese	–	–	698 (12.0)
Education (N, %)
Less than High School	1283 (23.7)	324 (15.3)	1045 (18.0)
High School Equivalent	2264 (41.8)	914 (43.3)	1068 (18.4)
Greater than High School	1866 (34.5)	874 (41.4)	3698 (63.6)
Lifestyle/Comorbidity Factors
Mean BMI, kg/m2 (SD)	28.1 (5.4)	28.5 (5.7)	28.4 (5.5)
Smoking Status (N, %)
Current Smoker	1147 (21.2)	128 (6.7)	762 (13.1)
Former Smoker	1995 (36.9)	978 (51.5)	2154 (37.1)
Never Smoked	2270 (41.9)	794 (41.8)	2896 (49.8)
Hypertension (N, %)	2052 (38.0)	1585 (75.7)	2590 (44.4)
Diabetes (N, %)	868 (16.1)	685 (33.2)	1572 (27.0)
Mean eGFR, mL/min/1.73 m2 (SD)	96.2 (13.7)	70.8 (17.7)	74.4 (16.6)
Mean Cholesterol, mg/dL (SD)	211.8 (40.9)	180.2 (41.2)	194.3 (35.9)

SD standard deviation, BMI body mass index, N number, MESA Multi-Ethnic Study of Atherosclerosis, ARIC Atherosclerosis Risk in Communities, eGFR estimated glomerular filtration rate.

Associations of PAA with cognitive function and dementia incidence

In fully adjusted model, each 5-year PAA at midlife was associated with lower late-life executive function (standardized difference: −0.14 [95% CI: −0.19, −0.09]) and global cognitive function (difference: −0.11 [95% CI: −0.16, −0.06]). There was no association between PAA and memory function. When using PAA at late-life, each 5-year PAA was cross-sectionally associated with lower memory (difference: −0.11 [95% CI: −0.18, −0.05]), executive function (difference: −0.19 [95% CI: −0.24, −0.14]) and global cognitive function (difference: −0.17 [95% CI: −0.23, −0.12]) (Table 3). The effect estimates of proteomics-based PAA on global cognitive function was comparable to that of APOE ε4 carriership. Specifically, individuals with one or two ε4 alleles had a 0.15 lower global cognitive function score compared to those with the ε3/ε3 genotype (95% CI: −0.20, −0.10). Similarly, each five-year increase in proteomics-based PAA was associated with a 0.17 lower cognitive function score, using the same set of covariates (Table S8).

Table 3.

Association between 5-year proteomics-based age acceleration at midlife and late-life and cognitive function Z-scores at late-life – ARIC midlife (1990–1992) and late-life (2011–2013)

	Midlife (N = 4783)		Late-life (N = 5123)
Cognitive Function Measures	Model 1a Difference (95% CI)	Model 2b Difference (95% CI)	Model 1a Difference (95% CI)	Model 2b Difference (95% CI)
Cognitive Domains
Memory	−0.08 (−0.14, −0.02)	−0.05 (−0.11, 0.02)	−0.17 (−0.23, −0.11)	−0.11 (−0.18, −0.05)
Executive Function	−0.17 (−0.22, −0.12)	−0.14 (−0.19, −0.09)	−0.27 (−0.32, −0.22)	−0.19 (−0.24, −0.14)
Global Cognitive Function	−0.14 (−0.20, −0.09)	−0.11 (−0.16, −0.06)	−0.25 (−0.30, −0.20)	−0.17 (−0.23, −0.12)

^aModel 1 is adjusted for chronological age at mid/late-life, sex, and race-center.

^bModel 2 is additionally adjusted for education, BMI smoking status, systolic blood pressure, diabetes status, cholesterol, and estimated glomerular filtration rate.

CI Confidence interval, ARIC Atherosclerosis Risk in Communities.

In the fully adjusted model, at ARIC midlife, each 5-year PAA was associated with 20% higher risk of incident dementia (HR: 1.20, 95%CI: 1.04, 1.36). Each 5-year PAA at ARIC late-life was more prominently associated with dementia risk with a HR of 2.14 [95%CI: 1.67, 2.73] (Fig. 2; the source data is available in S 4).

Fig. 2. Association between proteomics-based age acceleration per 5 years (5-year discrepancy between chronological and biological age) and dementia incidence.

Model 1 is adjusted for chronological age, sex, race/ethnicity, and study center. Model 2 additionally adjusted for education, BMI, smoking status, systolic blood pressure, diabetes status, cholesterol, and estimated glomerular filtration rate. Number of cases: ARIC midlife = 2251; ARIC late-life = 707; MESA = 506. Circles indicate hazard ratios and error bars represent 95% confidence intervals. PAA = proteomics-based age acceleration, MESA = Multi-Ethnic Study of Atherosclerosis, ARIC = Atherosclerosis Risk in Communities.

Validation in MESA

Similar to ARIC findings, in MESA, PAA was prospectively associated with lower cognitive function performance (Table 4). Similarly, each 5-year PAA (using ARIC midlife PAC) was associated with 1.23 [95%CI: 1.04, 1.46] higher hazard of dementia (Fig. 2). When using ARIC late-life PAC (proteins and coefficients based on ARIC late-life PAC) in MESA, we observed stronger effect estimates with dementia risk (1.61 [95%CI: 1.29, 2.01]) (Table S9) and similar results with cognitive function (Table S10).

Table 4.

Association between 5-year proteomics-based age acceleration at MESA Exam 1 (2000-2002) and cognitive function Z-scores at Exam 5 (2010–2011) - MESA

	MESA (N = 4,057)
Cognitive Function Measures	Model 1b Difference (95% CI)	Model 2c Difference (95% CI)
CASI Score	−0.13 (−0.19, −0.07)	−0.08 (−0.14, −0.03)
Digit Span Forward	−0.10 (−0.15, −0.04)	−0.06 (−0.12, −0.01)
Digit Span Backward	−0.14 (−0.20, −0.08)	−0.09 (−0.15, −0.04)
Digit Symbol Substitution	−0.14 (−0.19, −0.09)	−0.07 (−0.12, −0.03)

^aExam 1 proteomics-based age acceleration is calculated using proteins and coefficients from ARIC midlife.

^bModel 1 is adjusted for chronological age, sex, race/ethnicity, and study center.

^cModel 2 is additionally adjusted for education, BMI, smoking status, systolic blood pressure, diabetes status, cholesterol, and estimated glomerular filtration rate.

CASI The Cognitive Abilities Screening Instrument, CI Confidence interval, MESA Multi-Ethnic Study of Atherosclerosis.

Sensitivity analyses in the ARIC cohort

Using a non-imputed data did not change the association of proteomics-based PAA and cognitive function (Table S11). Using the selection of proteins and regression coefficients from ARIC midlife and protein levels from ARIC late-life, the effect estimates for late-life PAC lie between effect estimates from midlife and late-life (Table S12). Adjusting additionally for physical activity, healthy diet score, social isolation, traumatic brain injury and depression (in late-life only) attenuated the effect estimates but did not change the findings (Table S13). Stratifying by race (Table S14) APOE ε4 allele carriership (Table S15), sex (Table S16) didn’t change the findings.

Discussion

In this study, we show that higher PAA, which reflects the deviation of biological age from chronological age, is associated with lower performance in cognitive tests, particularly in relation to executive function and processing speed, and a higher risk of developing dementia. The associations were independent of chronological age, demographic, and cardiovascular risk factors. Our results indicate that PACs can be considered as a tool to identify individuals at risk for cognitive impairment and developing dementia in the future.

Prior studies investigated roles of different types of biological clocks in predicting future risk of cognitive impairment and dementia^38–44. For instance, various DNA methylation epigenetic clocks have been tested by multiple studies as marker for advanced cognitive aging and dementia incidence^45–48. Combining the data in a systematic review and meta-analysis, Zhou et al. showed that majority of these studies did not show a significant association and concluded that there is insufficient evidence to indicate that epigenetic aging can serve as a valid biomarker to individuals at risk for cognitive impairment and dementia⁴⁴. However, several well-powered longitudinal studies have consistently shown that accelerated aging or a faster pace of aging, as measured by epigenetic clocks such as PhenoAge, GrimAge, and DunedinPACE, is associated with more rapid cognitive decline and an increased risk of dementia^39,41–43. The reason for mixed and inconclusive results could be due to using a heterogenous group of DNA methylation aging clocks⁴⁹. Prior studies have shown that proteins have the potential to serve as metrics for quantifying biological aging^50,51. Proteins can be accurately measured and are closer to phenotypic expression³. Moreover, in clinical contexts, proteins are more useful, as medical professionals routinely rely on plasma proteins as biomarkers for diagnosing medical conditions, predicting outcomes, and assessing treatment efficacy. Although fewer in number, studies using proteomic data to assess PAA have similarly reported that proteomic biological aging markers are associated with an elevated risk of mild cognitive impairment, an incident of Alzheimer’s disease, and all-cause dementia^52,53. Sathyan et al. showed that a higher age acceleration, using proteomics clocks, predicts risk of motor cognitive risk syndrome, a pre-dementia syndrome characterized by slow gait and subjective cognitive concerns⁵⁴. Another study including UK Biobank individuals, showed that a higher proteomics aging clock is associated with a higher risk of all-cause dementia⁵³. In this study, we developed PACs in both midlife and late-life and demonstrated that both clocks, more strongly the late-life clock, predict dementia risk across two independent cohorts representing diverse racial and ethnic groups. Future studies with focus on application of these clocks in clinical settings for prediction and patient risk stratification are warranted to bring the scientific evidence closer to clinical practice.

Notably, we observed a stronger association with risk of dementia when using clocks developed in late-life as opposed to those at midlife. This finding might reflect the dynamic nature of biological markers across the lifespan and highlight the importance of considering age-specific changes in disease prediction models. The stronger predictive value of PACs in older age can be due to the fact that there is a greater variability in protein levels in older age, potentially making it a more effective tool for discerning differences. In addition, with aging there is a progressive accumulation of molecular alterations, such as increased oxidative stress, impaired protein clearance mechanisms, and chronic inflammation^3,9. These age-related changes may contribute to distinct proteomic signatures that can better reflect the evolving pathological processes underlying dementia development in later life⁵⁵. Another possibility is that late-life PACs are closer to the onset of dementia, making them potentially superior predictors. While proteomic alterations in midlife may reflect early pathological changes associated with dementia, they may not fully capture the complexity of the disease cascade that unfolds over several decades. To investigate whether the protein selection at older age is the driving factor, we constructed a PACs in older age using identified proteins from midlife, instead of those selected at late-life, then evaluated the association in the ARIC test set. While we observed a decline in the magnitude of effect estimates for dementia risk using this clock, the estimates remained stronger than midlife estimates, suggesting that the stronger late-life estimates are not solely due to the combination of proteins in the biological clock at late-life. As the effect estimate was still stronger than the midlife biological clock, it is possible that both factors mentioned earlier contribute to the difference between midlife and late-life estimates.

Dementia has a long preclinical phase, which typically takes decades to manifest as cognitive function impairments. To explore whether the PACs can be used to predict dementia risk at earlier stages of cognitive decline, we evaluated the association of the clocks with cognitive function. While both midlife and late-life clocks were associated with decrease in global cognition and executive function, only the late-life clock was associated with memory function. A possible explanation could be that usually impairment in executive function precedes memory impairments before full-blown dementia is presented^56,57.

Further exploration of the top overlapping individual proteins identified in both midlife and late-life revealed several plausible biological functions, including cell adhesion (Lumican [LUM], CUB domain-containing protein 1 [CDCP1], Coiled-coil domain-containing protein 80 [CCDC80], WNT1-inducible-signaling pathway protein 2 [CCN5], EGF-containing fibulin-like extracellular matrix protein 1 [EFEMP1], Scavenger receptor class F member 2 [SCARF2]), cell proliferation (Pleiotrophin [PTN], T Transgelin [TAGLN], Cartilage acidic protein 1 [CRTAC1]), neuronal development (Chordin-like protein 1 [CHRDL1], Neurogenic locus notch homolog protein 3 [NOTCH3]), and stress response and senescence (Transgelin [TAGLN], Growth/differentiation factor 15 [GDF15])^58–63. Several of these proteins have established roles in neurodegeneration or age-related cognitive decline. For instance, PTN is a secreted cell-signaling cytokine that acts as a modulator of multiple neuronal functions during development⁶⁴. In adults, PTN expression is limited to specific brain regions, including the cortex and hippocampus^64,65. In fact, a study using a proteomic approach identified elevated levels of endogenous PTN peptides in individuals with Alzheimer’s disease, suggesting its potential utility as a biomarker for diagnosis⁶⁶. Additionally, several of these proteins—PTN, CHRDL1, GDF15, and SCARF2—have been previously identified in studies as predictors of aging, further supporting the value of proteomics-based aging clocks as tools for risk stratification and early identification of individuals at risk for dementia in clinical and population settings.

Our study had several strengths, including a large sample size, representation of different racial and ethnic groups, external validation of the results in an independent cohort, multiple assessments of proteomics over time, longitudinal data collection spanning midlife and late-life, and availability of detailed information about patient characteristics and potential confounders. We also acknowledge several limitations of this study. First, information on dementia subtypes was not available in all participants, and there were differences in the methods for ascertaining dementia between the two cohorts. In addition, ARIC and MESA used different cognitive tests to assess global and domain-specific cognitive function. Nevertheless, we observed similar associations with both midlife and late-life clocks derived from the ARIC study in the MESA cohort, underscoring the robustness of our findings. Although we accounted for multiple demographic and cardiovascular factors in our analyses, data on other factors such as hearing, vision loss, and air pollution were not available. Therefore, we cannot rule out the possible effect of unmeasured confounders in the observed associations. Third, PACs are limited in identifying proteins responsible for dementia risk that aren’t age-related.

This study provides evidence regarding the utility of PACs for predicting dementia and cognitive impairment. The robust link between proteomic profiles and future dementia risk, particularly in late life, has the potential for translation in clinical practice for early detection of high-risk individuals and implementation of preventive strategies in individuals at risk.

Author contributions

S.S. and A.P. conceived and designed the study. S.S., A.P. T.M.H., J.C., J.S.P., R.D., R.D. J.I.R., P.G., and P.L.L., obtained funding. S.P., R.F.W., S.W., and J.L. conducted statistical analysis. S.S. wrote the first draft of the paper. T.M.H, B.S., W.T., J.C., K.A.W., R.C., A.C.W., and W.G. provided technical assistance. S.S., A.P., and W.G. supervised the study. All authors contributed to revisions of the paper and approved the final version as submitted.

Peer review

Peer review information

Communications Medicine thanks Ilja Demuth, Kazuhiko Uchida, and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

ARIC and MESA data are available through the NHLBI Biologic Specimen and Data Repository Information Coordinating Center (https://biolincc.nhlbi.nih.gov/studies/aric, https://biolincc.nhlbi.nih.gov/studies/mesa/). Additional requests for clinical or proteomic data from individual investigators may be submitted to the ARIC and MESA steering and publication committees and will be reviewed to ensure that data can be shared without compromising patient confidentiality or breaching intellectual property restrictions. Participant-level demographic, clinical and proteomic data may be partially restricted based on previously obtained participant consent. Data-sharing restrictions may also be applied to ensure consistency with confidentiality or privacy laws and considerations. The source data for Fig. 2 is in Supplementary Data 4.

Code availability

All analyses were performed using publicly available R software (R version 4.4.2 (https://www.r-project.org/), and SAS 9.4. The code used in this study can be made available from the corresponding author on reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01096-y.