Plasma-based brain age as a potential biomarker for cognitive health and risk of brain-related disorders

Abstract

Background

Understanding brain aging is essential for identifying early markers of cognitive decline. This study aimed to develop plasma-based biomarkers of brain aging and examine their associations with cognitive function.

Methods

We used data from 53,005 UK Biobank participants (2006–2010) with available Olink proteomics data (mean age 57 ± 8 years, 54% women). Protein levels were used to estimate brain age, organismal age, and conventional proteomic age using Least Absolute Shrinkage and Selection Operator (LASSO) regression. We assessed the relationships between these biological ages and cognitive function using multivariable-adjusted linear regression models, and their association with incident Alzheimer’s disease (AD) and stroke using Cox proportional hazards models. Findings were validated in 2066 participants from the Framingham Heart Study (FHS) Offspring cohort (Exam 8, 2002–2005, mean age 67 ± 9 years, 55% women), also using the Olink platform.

Results

Accelerated brain aging is significantly associated with poorer cognitive performance, whereas organismal and conventional proteomic ages are not. All three biological age measures are linked to increased risk of AD and stroke, but brain aging shows the strongest association (hazard ratio (HR) for AD: 1.79 [95% confidence interval (CI): 1.66–1.93]; HR for stroke: 1.25 [95% CI: 1.17–1.33]). In the FHS validation, brain aging is associated with lower performance in cognitive domains such as attention and visual memory, and with increased risk of AD (HR: 1.64 [95% CI: 1.37–1.97]).

Conclusions

Plasma-based biomarkers of brain aging may offer a promising tool for monitoring cognitive health and identifying individuals at risk for age-related diseases.

Plain language summary

This study looked at blood proteins to find signs of brain aging and their link to cognitive function. Researchers used protein data from over 53,000 people in the UK Biobank and a machine learning technique to create different “biological age” scores. The brain aging score was strongly tied to lower cognitive performance and a higher risk of Alzheimer’s disease and stroke. The findings were tested in the Framingham Heart Study and showed similar results—older brain age was linked to worse memory and a higher risk of Alzheimer’s disease. This suggests that blood tests focused on brain aging could help doctors spot early signs of cognitive decline and act sooner to prevent age-related diseases.

Wang et al. develop plasma-based biomarkers of brain aging using proteomic data from the UK Biobank and validate their findings in the Framingham Heart Study. Acceleration in plasma-based brain aging is linked to poorer cognitive performance and shows a positive association with Alzheimer’s disease and all-cause mortality.

Introduction

As the global population ages, research on brain aging is of paramount importance, with increasing numbers of individuals at risk for chronic conditions such as Alzheimer’s disease (AD), and other neurodegenerative diseases¹. Understanding the biological processes that contribute to brain aging and cognitive decline is critical for developing early diagnostic tools and biomarkers, effective interventions, and potential cures².

Brain aging can be evaluated using various methods, including neuroimaging techniques, cognitive assessments, blood-based or cerebrospinal fluid biomarkers, and cellular or molecular aging markers, such as telomere length and epigenetic clocks. Early detection often relies on laboratory tests; for instance, biomarker divergence (e.g., amyloid-beta) between AD and cognitively normal populations occurs approximately 12 years before cognitive decline becomes apparent in AD diagnosis³. In addition to this hallmark of AD, biological age estimates derived from plasma proteins (e.g., the proteomic age clock) can also serve as predictors of AD^4–6. Identifying aging-related proteins may provide valuable insights into the early diagnosis of AD.

The identification of aging-related proteins is also crucial for gaining a deeper understanding of the brain aging process. Proteins and their coding genes provide insights into the biological mechanisms underlying brain aging^7,8. For instance, beyond the genetic mapping of AD, proteomic studies not only validate well-established amyloid and tau pathways but also uncover additional components within broader protein networks. These include processes such as RNA splicing, neurodevelopment, immune response, membrane transport, lipid metabolism, synaptic function, and mitochondrial activity⁹. Developing precise proteomic aging models for the brain can help identify proteins most linked to susceptibility to cognitive decline, enabling the development of targeted early interventions and monitoring the effectiveness of treatments^10,11.

Several large population studies (e.g., UK Biobank, UKB) have demonstrated that the proteomic age clock is a reliable biomarker for predicting mortality and morbidity^4,5,12. However, traditional approaches to estimating the proteomic age clock do not account for variations in aging rates across different organs and tissues, thereby overlooking aging heterogeneity throughout the whole body¹³. Recently, organ-specific proteomic age clocks have been introduced to refine aging estimation at the organ level. While conventional proteomic aging models may exhibit stronger overall associations with various diseases, organ-specific aging models often outperform them in predicting diseases specific to particular organs^14–17. For instance, brain-specific aging proteins show stronger associations with AD and related dementias compared to proteins associated with other organs. Beyond disease outcomes, the association between brain-specific aging proteins and early signs of AD (e.g., memory decline) remains underexplored, as does the reproducibility and utility of brain-specific aging models for cognitive function assessment.

The goal of this project is to develop plasma-based brain aging biomarkers using data from the UKB and to validate our findings in the Framingham Heart Study (FHS). Additionally, we examine the association between brain-specific aging proteins with cognitive function and the risk of age-related conditions in these two prospective cohorts. We observe that increasing plasma-based brain age is associated with poorer cognitive performance—particularly in attention/concentration and memory—and with a higher risk of AD. These associations between plasma-based brain age and cognition or AD are more significant than those observed with conventional proteomic age. Our findings highlight the potential of plasma-based brain age as a promising tool for monitoring cognitive health and identifying individuals at risk for age-related diseases.

Methods

Study population

The UKB is a large-scale prospective cohort study that includes comprehensive genetic and phenotypic data from 502,505 participants residing in the United Kingdom, recruited between 2006 and 2010^18,19. For this analysis, we focused on a subset of the UKB cohort, specifically participants with baseline Olink Explore proteomics data available, resulting in a study sample of 53,014 individuals⁷.

The Framingham Heart Study (FHS), initiated in 1948, is a well-established community-based, prospective cohort study²⁰. In 1971, the Offspring cohort was recruited, which consists of the children of the original participants along with their spouses²¹. The Offspring cohort has undergone ten rounds of health examinations, typically conducted every four to six years, beginning with the first cycle from 1971 to 1975. For this study, we analyzed data from 2066 individuals in the Offspring cohort who participated in the eighth examination cycle (2005–2008), during which blood samples were collected with the purpose of proteomics profiling.

The UK Biobank study was conducted under generic ethical approval as a research tissue bank from the NHS North West Research Ethics Committee (21/NW/0157). This approval covers the present research, and no additional ethical approval was required. Access to the UK Biobank resource is open to bona fide researchers undertaking health-related research in the public interest (www.ukbiobank.ac.uk/register-apply/). The present study was conducted using UK Biobank data under application number 76269. The FHS protocol was approved by the Institutional Review Board of the Boston University Medical Center. All FHS participants provided written informed consent for the collection and use of their data and biological samples.

Proteomics profiling

In the UKB study, baseline proteomics data were collected from participants during visits conducted between 2006 and 2010. Proteomic analysis involved measuring 2923 proteins using the Olink Explore platform, which provides Normalized Protein Expression (NPX) in log2 scale values for each protein. To ensure data quality and reliability, twelve proteins with a missing data rate exceeding 20% across all samples were excluded, resulting in the retention of 2911 proteins for further analysis. For the remaining proteins with missing values, imputation was performed by replacing missing values with the mean NPX values calculated from non-missing samples⁸.

For the FHS, the methodology for proteomics profiling has been extensively documented in the manufacturer website (https://olink.com/knowledge/documents). Briefly, plasma samples were collected during research visits and stored at −80 °C until analysis. Protein concentrations were measured using the Olink Explore HT platform. The protein concentration was also measured as NPX in log2 scale. A total of 5,371 proteins were analyzed. Of these, 24 proteins were excluded due to being control probes, 135 proteins were removed because we could not find them in the Gene Tissue Expression Atlas (GTEx). The remaining 5212 proteins (their missing rate<20%) with missing values were imputed by the means of other non-missing samples. Additionally, the 2750 overlapped and quality-controlled proteins among UKB and FHS were selected for downstream analysis (Supplementary Fig. 1).

Organ-specific proteins assignment

RNA sequencing data from the GTEx v8 were used to analyze gene expression across various tissues²². The tissues were grouped according to their respective organs (Supplementary Data 1), and RNA-seq read counts were normalized using DESeq2²³. For each tissue, mean normalized read counts were calculated across all samples. Organ-level expression was represented by the maximum normalized mean among the tissues within each organ. To identify organ-enriched genes, a fold-change was computed as the ratio of the maximum expression level in one organ to the second-highest expression level in another organ. Genes were considered organ-enriched if the fold-change exceeded four²⁴. A list of GTEx genes and their corresponding enriched organs was generated (Supplementary Data 2). Subsequently, the quality-controlled proteins from the UKB and FHS Olink dataset were mapped to the GTEx genes and assigned to specific organs for further analysis. Plasma-based brain age proteins were defined as those with coding genes enriched in the brain (genes mapped to the brain with enrich_organ_yes = 1 in Supplementary Data 2). Organismal proteins were defined as those with coding genes expressing across multiple organs (enrich_organ_yes = 0 in Supplementary Data 2). Conventional age proteins were those available and passed quality control in both UKB and FHS. No proteins were found to be enriched in multiple organs (Supplementary Data 3, Supplementary Data 4). Among the 2,750 overlapped proteins (i.e., conventional age proteins), there were 102 (3.7%) plasma-based brain age proteins and 2,235 (81.3%) organismal age proteins (Supplementary Fig. 1).

Cognitive and other health outcomes

Participants in the UKB completed their cognitive tests either in-person in the assessment centers or during the online follow-up. Six cognitive tests were used to assess different domains^25,26: Fluid intelligence (FI) assessed for verbal and numerical reasoning, Numeric memory (MEMN) for working memory, Pairs matching (MEMP) for visual declarative memory, Symbol digit substitution (SDS) for processing speed, Reaction time (RT) for processing speed, and Trail making test B (TMT B) for executive function. We developed a general cognitive score, which was the first unrotated principal component generated by the principal component analysis of the previous six normalized cognitive test scores²⁷. The scores of some tests were transformed as shown elsewhere²⁸. All the cognitive tests were performed simultaneously with the lab tests performed to measure proteins (UKB baseline, year 2006 to 2010). Higher cognitive scores meant better/higher cognitive function.

All-cause mortality and six common chronic conditions (i.e., diabetes, heart failure, myocardial infarction, atrial fibrillation, stroke, and AD) were defined using specific UKB Field IDs (Supplementary Data 5). Outcome selection is based on data availability and scientific rationale. Because the FHS was originally designed as a community-based cohort for cardiovascular research, incident cardiovascular events are comprehensively ascertained and well powered. AD, stroke, and all-cause mortality were included for their hypothesized association with plasma-based brain age, while several heart-related outcomes were added as negative controls because they are less likely to be directly linked to plasma-based brain age. This design allowed us not only to evaluate plasma-based brain age with brain-related outcomes but also to test the specificity of brain age as a potential biomarker.

The cognitive function in FHS Offspring was evaluated using a battery of neuropsychological (NP) tests that target multiple cognitive domains²⁹. Details of every NP test were described elsewhere³⁰. The domain of verbal memory was determined/assessed by the Wechsler Memory Scale-III Logical Memory (LM) tests as an average score of immediate recall, delayed recall, and recognition tests (i.e., average score of LMi, LMd, LMr). Verbal learning was obtained by the average of Paired associate learning (PAS) immediate recall and delayed recall tests (i.e., average score of PASi, PASd), and visual memory was obtained by the average of several visual reproduction (VR) tests (i.e., average score of VRi, VRd, and VRr). Trail making test A (TMT A) and TMT B were used to assess attention and concentration: a natural log transformation was applied to normalize the distributions of the TMT A and TMT B. To maintain interpretative consistency, the transformed values were further re-signed so that higher TMT A or TMT B correspond to better task performance^31,32. The remaining NP tests (WAIS-III Similarities subtest, Boston Naming Test (30-item version), and Hooper Visual Organization Test) each represented a unique cognitive domain (abstract reasoning, language, and visuoperceptual organization)³². For those who have multiple times of NP tests, we used the test that was closest to the time when lab tests were performed and proteins measured. Higher NP test scores reflected better cognitive performance.

Adverse events in FHS, including deaths, congestive heart failure, myocardial infarction, atrial fibrillation, and stroke, were adjudicated by a panel of 3 investigators using pre-established criteria³³. Diabetes was diagnosed either by a fasting plasma glucose level ≥ 126 mg/dL or by treatment with insulin or an oral hypoglycemic agent³⁴. AD diagnosis was ascertained based on available data from NP test results, neurological examination, a family interview, FHS cycle exam records, and/or hospital or nursing home medical records reviewed by a panel of at least 1 neurologist and 1 neuropsychologist following standard research criteria³⁵.

Statistics and Reproducibility

Organ aging model

In this study, a random sampling approach was used to split the UKB data into a training set (70%, n = 37,105) and a testing set (30%, n = 15,900). Least absolute shrinkage and selection operator (LASSO) regression models were fitted in the training dataset³⁶, with chronological age as the outcome and each protein set (brain, organismal, or conventional) along with sex as predictors. Hyperparameter tuning was performed using five-fold cross-validation to select the optimal regularization parameter (λ) that minimized the mean squared error (MSE). The final model was derived from 500 bootstrap iterations, and the mean predicted age across these iterations was calculated. The coefficients obtained from the training set were then applied to the testing dataset, where the mean predicted age was similarly computed across the 500 bootstrap models.

Functional enrichment analyses of organ-specific aging-related proteins

Functional enrichment analyses were performed using the g:Profiler web server (https://biit.cs.ut.ee/gprofiler/gost) for the top proteins selected for different age³⁷. The analyses were conducted under the g:GOSt module with default query settings, examining pathways in Gene Ontology (GO) (biological process, molecular function, cellular component), Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and WikiPathways. Multiple testing for enrichment P values (P-adjusted) was corrected using the g:SCS algorithm, which accounts for the uneven distribution of functionally annotated gene sets³⁸.

Calculate the individual organ age gap

A linear regression was fitted between the predicted organ age and chronological age to evaluate their relationship. The age gap was defined as the residual difference between predicted organ age and the linear regression estimate based on chronological age^17,39. This calculation was performed separately for both the training and testing datasets.

Association of organ age gaps and outcomes

While the training/testing split was maintained for age-model development to minimize overfitting in organ-age estimation, the sets were merged for downstream association analyses to maximize sample size and enhance statistical power (n = 53,004). Linear regression models were developed to examine the association between age gaps and each standardized cognitive score. Models were adjusted for age, sex, smoking status, body mass index (BMI), physical activity, diet, alcohol consumption, years of schooling, hypertension, and APOE4 allele dosage. More details of these covariates were provided in a previous study⁴⁰. Cox proportional hazards models were used to assess the relationship between organ age gaps and the development of incident outcomes, excluding prevalent cases from the analysis. Similar covariates as with the linear regression models were included in the Cox models. We also applied a time-stratified Cox model, in which strata were split at the median follow-up time when the proportional hazards assumption was violated. In the secondary analysis, participants were further divided into three groups: normal aging, defined as predicted age gaps within two standard deviations of the population mean; accelerated aging, defined as predicted age gaps with more than two standard deviations above the mean; and decelerated aging, defined as predicted age gaps with more than two standard deviations below the mean. Multivariable-adjusted Kaplan-Meier (KM) curves were used to illustrate the cumulative risk of the selected outcomes (i.e., all-cause mortality, stroke, and AD) across different aging groups during the follow-up period.

Association of organ age gaps and outcomes in the FHS

External validation of the organ age models was performed in the FHS Offspring Cohort. Organ age in the FHS was calculated using the regression coefficients from UKB-trained LASSO model. Because the FHS participants (mean age = 67 years) were on average 10 years older than those in the UKB (mean age = 57 years), we recalibrated the models by adjusting the intercept term in the LASSO regression to better align the predicted and actual ages in the FHS⁴¹.

Reproducibility of the organ aging models was further evaluated in the FHS as a secondary analysis. Using the full FHS sample (n = 2066), we applied a similar analytical strategy as in the UKB, including organ age modeling and age gap calculation, with a 70:30 training/testing split. LASSO regression models were fitted with chronological age as the outcome, using five-fold cross-validation for hyperparameter tuning and 500 bootstrap iterations in the training set. Organ-specific age was defined as the mean predicted age across the 500 bootstraps. The age gap was then calculated as the residual of organ age from chronological age and was estimated separately in the training and testing sets.

To validate the association between age gaps and NP test domains, we used linear regression models adjusted for age at NP test, sex, smoking, BMI, education year, hypertension, and APOE4 allele dosage. To validate the association between age gaps and the outcomes of interest, we used the multivariable adjusted Cox proportional hazards model adjusted for age of proteomics, sex, smoking, BMI, hypertension, and education year.

The analytic framework in this study is shown in Fig. 1. All the analyses were performed using R software package version 4.2.1 (https://www.r-project.org/) and Python version 3.6.8 (https://www.python.org/).

Fig. 1. Analytic framework of brain-specific age and health outcomes.

The overall study design consists of three parts: a Identify organ-specific proteins. We mapped proteins with corresponding genes and designated each protein as organ-specific or not according to whether its gene’s expression level was four-fold higher in one organ than any other organ in the GTEX v8 RNA sequencing data. There were a total of 2750 conventional proteins overlapped in the UK Biobank (UKB) and the Framingham Heart Study (FHS). Of those, 2235 were organismal proteins, while 102 were plasma-based brain proteins. b Use plasma proteins to predict chronological age. We separated the UKB into training and testing datasets (7:3 ratio). In the training set (N = 37,105), the 500 bootstrap aggregated LASSO regression models with five-fold cross-validation were fitted for the plasma-based brain proteins well as organismal and conventional proteins. Organ ages were the mean predicted age across 500 bootstraps. The age gaps were the residual of organ ages and the linear regression estimates based on chronological age. The LASSO models were applied to the UKB testing set (N = 15,900) to estimate the organ age gaps. External validation and replication have been conducted in the FHS Offspring Exam 8 cohort (N = 2066), the UKB trained LASSO models were recalibrated in the FHS to estimate the age gaps as the external validation; while similar analytic strategies (i.e., ratio of 7:3 on training/testing separation, 500 bootstrapping LASSO models) as the UKB were also implemented in the FHS as the replication analysis. c Associate age gaps with health outcomes. Both in the UKB (a combined set of training and testing) and the FHS, we used linear regression models to see the associations of organ aging and cognitive function and fitted Cox proportional hazard models to test the associations of organ aging with mortality and other outcomes. Figure created with BioRender.com.

Results

Cohort descriptive

Table 1 depicts the characteristics of the participants from the UKB and the FHS, respectively. In the UKB, participants with baseline proteomics data were randomly divided into training and testing sets in a 7:3 ratio. The UKB cohort had a mean age of 57 years, with women comprising 54% of the sample. While FHS participants had a mean age of 67 years, with 55.3% were women. The proportion of participants who had anti-hypertension medication was higher in the FHS compared to the UKB (51% vs. 22%).

Table 1.

Baseline characteristics of the study participants

Characteristics	UKB Training (n = 37,105)	UKB Testing (n = 15,900)	FHS Offspring (n = 2066)
Chronological age, years	57 ± 8	57 ± 8	67 ± 9
Female (%)	19,956 (53.8)	8621 (54.2)	1143 (55.3)
BMI, kg/m2	28 ± 5	27 ± 5	28 ± 5
Education years	14 ± 5	14 ± 5	14 ± 3
SBP, mmHg	138 ± 19	138 ± 19	129 ± 17
DBP, mmHg	82 ± 10	82 ± 10	73 ± 10
Anti-hypertension medication (%)	8269 (22.3)	3481 (21.9)	1042 (50.5)
Current smoking (%)	3963 (10.7)	1637 (10.3)	182 (8.8)
Excessive alcohol consumption (%)	9604 (25.9)	4139 (26.0)	-
Lack of physical activity (%)	4378 (11.8)	1865 (11.7)	-
Poor diet (%)	21,928 (59.1)	9398 (59.1)	-

Values are represented as n (%) for dichotomous variables or mean ± standard deviation (SD) for continuous variables. BMI Body mass index, SBP systolic blood pressure, DBP diastolic blood pressure.

Excessive alcohol consumption was defined as pure alcohol consumed more than 14 grams per day for women and more than 28 grams per day for men.

Lack of physical activity meant less than 150 minutes moderate intensity physical activity or 75 minutes of vigorous activity per week (or an equivalent combination).

Poor diet meant cannot meet at least 4 of the following 7 food groups diet recommendation: Fruits ≥ 3 servings/day, Vegetables ≥ 3 servings/day, Fish ≥ 2 servings/week, Processed meats ≤ 1 serving/week, Unprocessed red meats ≤ 1.5 servings/week, Whole grains ≥ 3 servings/day, and Refined grains ≤ 1.5 servings/day.

Correlations of chronological age and predicted organ age

A total of 2750 UKB and FHS overlapping proteins were included in the “conventional” aging model. Of these, 2235 were incorporated into the “organismal” aging model, while the remaining proteins were organ-specific. In the UKB, correlations between chronological age and both conventional and organismal aging models exceeded 0.9. Correlations between chronological age and the three aging measures are detailed in Supplementary Data 6. The root mean squared error of predicting chronological age by the conventional aging model was the lowest (as 2.90 and 3.07 in the UKB training and testing, respectively). while the brain aging model was the highest (as 5.75 and 5.81 in the UKB) among the three aging models (brain-specific, organismal, and conventional). Similarly, the three aging models were recalibrated to the FHS cohort, where their correlations with chronological age ranged from moderate to high (Pearson r = 0.68–0.90; Supplementary Data 6). Additionally, the proteins selected in at least 90% (≥450 times) of the 500 bootstrapped LASSO iterations in the UKB training set are listed in Supplementary Data 7.

Variation and correlation of organ-specific age gaps

Age gaps were calculated as the residuals from a simple linear regression of organ-specific ages on chronological age. While the mean age gaps were approximately zero across all models, their variability differed. The brain-specific age gap exhibited a higher standard deviation ( ~ 4 years), whereas the organismal and conventional age gaps had lower standard deviations (below 3 years). Additionally, there was a strong correlation between conventional and organismal age gaps (Pearson’s r = 0.91) (Fig. 2). The raw age gap, defined as the difference of predicted age minus chronological age, and its distribution have been shown in Supplementary Fig. 2.

Fig. 2. Correlations of organ age gaps and cognitive function in the UK Biobank (N = 53,005).

The Scatterplots below the diagonal show the relationship between each pair of age gaps and cognitive function, the Histograms on the diagonal represent the distribution of each variable, Pearson correlation coefficients between each pair of variables are shown above the diagonal. The organ age gaps showed positive correlations with one another, whereas they were all inversely correlated with each score of cognitive function. Cognitive function included fluid intelligence (FI), numeric memory (MEMN), pairs matching (MEMP), symbol digit substitution (SDS), reaction time (RT), trail making test B (TMT B), and general cognitive score (General cog).

Association of brain aging and cognitive function

Fig. 2 illustrates the correlation between each organ aging measurements and cognitive test scores. As shown in Table 2, plasma-based brain aging was inversely associated with all six cognitive tests after multiple testing correction, with the most significant associations observed for Reaction Time (effect size = −0.04, 95% CI = −0.05 to −0.02, P = 2.54 × 10⁻¹⁵), Numeric Memory (effect size = −0.04, 95% CI = −0.05 to −0.02, P = 3.12 × 10⁻⁵), and the Symbol Digit Substitution Test (effect size = −0.04, 95% CI = −0.05 to −0.02, P = 9.06 × 10⁻⁵). Additionally, brain aging was associated with the general cognitive score (effect size = −0.03, 95% CI = −0.05 to −0.01, P = 0.002). In contrast, organismal aging was only associated with Reaction Time (effect size = −0.03, 95% CI = = −0.03 to −0.02, P = 5.95 × 10⁻⁸) after multiple testing correction. Conventional age was associated with both Reaction Time (effect size = −0.02, 95% CI = −0.03 to −0.02, P = 7.17 × 10⁻⁸) and the Symbol Digit Substitution Test (effect size = −0.03, 95% CI = −0.04 to −0.01, P = 0.003) following correction for multiple testing.

Table 2.

Association of aging measurements and cognitive function in the UK Biobank (N = 53,005)

Cognitive function	Domain	Plasma-based brain age		Organismal age		Conventional age
		Effect size* (95% CI)	P value	Effect size* (95% CI)	P value	Effect size* (95% CI)	P value
Fluid intelligence, FI	Verbal and numerical reasoning	−0.02 (−0.03 to −0.01)	0.002	−0.00 (−0.02 to 0.01)	0.612	−0.00 (−0.02 to 0.01)	0.551
Numeric memory, MEMN	Working memory	−0.04 (−0.05 to −0.02)	3.12 × 10−5	−0.01 (−0.02 to 0.01)	0.467	−0.00 (−0.02 to 0.01)	0.773
Pairs matching, MEMP	Visual declarative memory	−0.02 (−0.03 to −0.01)	2.63 × 10−4	−0.00 (−0.01 to 0.00)	0.344	−0.01 (−0.02 to −0.00)	0.025
Symbol digit substitution, SDS	Processing speed	−0.04 (−0.05 to −0.02)	9.06 × 10−5	−0.02 (−0.03 to 0.00)	0.089	−0.03 (−0.04 to −0.01)	0.003
Reaction time, RT	Processing speed	−0.04 (−0.05 to −0.03)	2.54 × 10−15	−0.03 (−0.03 to −0.02)	5.95 × 10−8	−0.02 (−0.03 to −0.02)	7.17 × 10−8
Trail making B, TMT B	Executive function	−0.03 (−0.05 to −0.01)	0.006	−0.01 (−0.03 to 0.01)	0.235	−0.01 (−0.03 to 0.01)	0.156
General cognitive score	-	−0.03 (−0.05 to −0.01)	0.002	−0.01 (−0.03 to 0.00)	0.141	−0.02 (−0.04 to 0.00)	0.071

The multivariable linear regression models were adjusted for age, sex, smoking, body mass index (BMI), physical activity, diet, alcohol consumption, education year, hypertension, and APOE4 allele dosages.

*Expressed as per standard deviation. Bold font represents a significant association after Bonferroni correction (two-sided P < 0.05/7 = 0.007). Italic font represents nominally significant association (two-sided P < 0.05).

Association of brain aging with health outcomes

As depicted in Table 3, plasma-based brain aging was significantly associated with all selected outcomes, with the strongest associations observed for all-cause mortality and AD. A one-standard-deviation increase in brain aging corresponded to a 50% increase in all-cause mortality risk (HR = 1.50, 95% CI: 1.46–1.55) and an 79% increase in AD risk (HR = 1.79, 95% CI: 1.66–1.93). Similarly, organismal and conventional aging were associated with all the selected outcomes, though the associations were slightly attenuated. Sensitivity analysis of time-stratified Cox models remained highly consistent with the associations observed by the original Cox models (Supplementary Data 8).

Table 3.

Association of aging gaps and risk of incident adverse outcomes in the UK Biobank (N = 53,005)

Outcomes	#Events	Plasma-based brain age		Organismal age		Conventional age
		HR* (95% CI)	P value	HR* (95% CI)	P value	HR* (95% CI)	P value
All-cause mortality	4010	1.50 (1.46–1.55)	1.04 × 10−164	1.28 (1.24–1.32)	3.37 × 10−55	1.33 (1.29–1.38)	2.31 × 10-74
Diabetes	1936	1.23 (1.18–1.29)	2.15 × 10−21	1.08 (1.04–1.13)	4.04 × 10-4	1.08 (1.04–1.13)	3.99 × 10-4
Heart Failure	1626	1.40 (1.33–1.46)	9.25 × 10-45	1.34 (1.27–1.40)	3.02 × 10-32	1.35 (1.28–1.41)	4.26 × 10-33
Myocardial infarction	1131	1.22 (1.16–1.30)	4.24 × 10-12	1.21 (1.14–1.28)	1.59 × 10-10	1.23 (1.16–1.30)	7.29 × 10-12
Atrial fibrillation	2863	1.13 (1.09–1.18)	3.84×10-11	1.18 (1.14–1.23)	5.73 × 10-19	1.18 (1.14–1.22)	2.17 × 10-18
Stroke	1030	1.25 (1.17–1.33)	2.90 × 10-13	1.15 (1.09–1.23)	4.63 × 10-6	1.20 (1.12–1.27)	1.23 × 10-8
Alzheimer’s disease	533	1.79 (1.66–1.93)	5.82 × 10-50	1.28 (1.17–1.39)	2.09 × 10-8	1.46 (1.34–1.59)	8.51 × 10-19

Multivariable cox proportional hazard models were adjusted for age, sex, smoking, BMI body mass index, physical activity, diet, alcohol consumption, education years, hypertension, and APOE4 risk allele dosages.

*Expressed as per standard deviation. All associations were significant and represented by bold after Bonferroni correction (two-sided P < 0.05/7 = 0.007).

To further examine these relationships, we categorized participants into accelerated aging, normal aging, and decelerated aging groups. Compared to normal aging, accelerated brain aging was associated with a higher risk of all selected outcomes (Fig. 3a, Supplementary Data 9), whereas decelerated brain aging was linked to a lower risk (Fig. 3b, Supplementary Data 9). Similar trends were observed for organismal and conventional aging, though with lower hazard ratios.

Fig. 3. Incident risk of adverse events of extreme aging versus normal aging in the UK Biobank (N = 53,005).

a The hazard ratio (HR) of accelerated aging vs. normal aging, b the HR of decelerated aging vs. normal aging. The x-axis of the bubble plot displays the HRs for extreme aging groups compared to normal aging group by Cox model adjusted for age, sex, smoking, BMI, physical activity, diet, alcohol consumption, education years, hypertension, and APOE4 allele dosages. The size of the bubble represents significance level (i.e., -log10(P)). Name of the proteomic age besides the bubble indicating significant associations between the proteomic age and the adverse event after Bonferroni correction (two-sided P < 0.05/7 = 0.007). Detailed effect sizes and P values (two-sided) are shown in Supplementary Data 9.

To assess the cumulative risk (i.e., absolute risk) of all-cause mortality and brain-related diseases, we used the multivariable adjusted KM plots to show the probability of an adverse event across follow-up time among extreme aging and normal aging groups (Fig. 4). The long-term (e.g., 15 years) probability of death, developing AD, or developing stroke were significantly higher in the accelerated aging group as compared to normal or decelerated aging groups in brain, organismal, and conventional aging models (log-rank test P value all less than 0.05). The long-term probability of death was as high as 30% for those who had accelerated brain aging, and their chances of developing stroke or AD in 15 years were all above 5% as compared to around 4% for those who were in accelerated aging of organismal or conventional groups.

Fig. 4. Multivariable adjusted Kaplan-Meier plots among accelerated, decelerated, and normal aging groups in the UK Biobank (N = 53,005).

a–i Were the cumulative risks of all-cause mortality, stroke, and AD by plasma-based brain, organismal, and conventional aging groups respectively. The multivariable Kaplan-Meier (KM) plots were adjusted for age, sex, smoking, BMI, physical activity, diet, alcohol consumption, education years, and APOE4 allele dosages. P values were calculated by the log-rank test (two-sided).

Validation and Replication in the FHS

We further validated our findings in the FHS. Table 4 displays the association of brain aging with different NP tests. Most of the NP tests were associated with brain aging after multiple testing correction, except the Boston Naming Test. The most significant association was observed between brain aging and attention and concentration, which was tested by Trail Making Test A and B (multivariable-adjusted effect size = −0.10, 95% CI = −0.14 to −0.06, P = 8.77 × 10⁻⁸). Attention and concentration was also nominally associated with organismal aging, and remained significantly associated with conventional aging after multiple testing correction (effect size = −0.07, 95% CI = −0.11 to −0.04, P = 8.58 × 10⁻⁵).

Table 4.

Association of aging gap and neuropsychological (NP) test domains in the FHS (N = 2066)

NP test	Domain	Plasma-based brain age		Organismal age		Conventional age
		Effect size* (95% CI)	P value	Effect size* (95% CI)	P value	Effect size* (95% CI)	P value
Logical Memory Tests	Verbal memory	−0.06 (−0.10 to −0.02)	0.004	0.01 (−0.03 to 0.05)	0.6	−0.01 (−0.05 to 0.03)	0.69
Visual Reproduction Tests	Visual memory	−0.08 (−0.12 to −0.05)	9.45 × 10-6	−0.03 (−0.06 to 0.01)	0.18	−0.03 (−0.07 to 0.00)	0.08
Paired Associate Learning Tests	Verbal learning	−0.09 (−0.14 to −0.05)	2.06 × 10-5	−0.04 (−0.08 to 0.00)	0.08	−0.05 (−0.09 to −0.01)	0.03
Trail Making Test A and Test B	Attention and concentration	−0.10 (−0.14 to −0.06)	8.77 × 10-8	−0.05 (−0.09 to −0.02)	0.004	−0.07 (−0.11 to −0.04)	8.58×10-5
Similarities subtest	Abstract reasoning	−0.09 (−0.13 to −0.05)	5.37 × 10-5	−0.02 (−0.07 to 0.02)	0.3	−0.04 (−0.09 to 0.00)	0.06
Boston Naming Test	Language	−0.06 (−0.10 to −0.01)	0.009	0.00 (−0.04 to 0.05)	0.83	−0.02 (−0.07 to 0.02)	0.32
Hooper Visual organization Test	Visuoperceptual organization	−0.10 (−0.15 to −0.06)	9.45 × 10-6	−0.02 (−0.06 to 0.02)	0.36	−0.03 (−0.07 to 0.02)	0.21

The multivariable linear regression models were adjusted for age at cognitive testing, sex, smoking, body mass index (BMI), years of completed education, prevalent hypertension, and APOE4 allele dosages.

*Expressed as per standard deviation. Bold font represents significant association after Bonferroni correction (two-sided P < 0.05/7 = 0.007). Italic fond represents nominal significant association (two-sided P < 0.05).

As shown in Table 5, plasma-based brain age in the FHS was positively and significantly associated with all-cause mortality and AD. A one-standard-deviation increase in brain aging corresponded to a 45% increase in all-cause mortality risk (HR = 1.45, 95% CI: 1.31–1.58), and 64% increase in AD risk (HR = 1.64, 95% CI: 1.37–1.97). Similarly, conventional aging was linked to the same two outcomes as well as heart failure (HR = 1.41, 95% CI: 1.15-1.73), while conventional aging was associated with mortality and heart failure but not AD (HR = 1.13, 95% CI: 0.95–1.34).

Table 5.

Association of aging gaps and risk of incident outcomes in the FHS Offspring (N = 2066)

Outcomes	#Events	Plasma-based brain age			Organismal age		Conventional age
		HR* (95% CI)	P value	HR* (95% CI)		P value	HR* (95% CI)	P value
All-cause mortality	553	1.45 (1.34–1.58)	2.84 × 10-18	1.19 (1.09-1.29)		4.29 × 10-5	1.19 (1.09–1.29)	4.45 × 10-8
Diabetes	103	0.92 (0.76–1.13)	0.45	0.97 (0.80–1.18)		0.77	0.97 (0.80–1.19)	0.80
Heart Failure	89	1.27 (1.02–1.59)	0.03	1.36 (1.12–1.66)		0.002	1.41 (1.15–1.73)	0.001
Myocardial infarction	60	1.21 (0.93–1.56)	0.16	1.09 (0.85–1.39)		0.51	1.17 (0.91–1.51)	0.22
Atrial fibrillation	209	1.10 (0.96–1.26)	0.19	1.21(1.06–1.38)		0.006	1.19 (1.04–1.36)	0.01
Stroke	80	1.08 (0.86–1.35)	0.51	0.98 (0.78–1.22)		0.83	1.00 (0.80–1.25)	0.99
Alzheimer’s disease	128	1.64 (1.37–1.97)	5.90 × 10-8	1.13 (0.95–1.34)		0.16	1.30 (1.09–1.54)	0.003

Multivariable cox proportional hazard models were adjusted for age, sex, smoking, body mass index (BMI), hypertension, education year, and APOE4 risk allele dosages.

*Expressed as per standard deviation. Bold font represents significant association after Bonferroni correction (two-sided P < 0.05/7 = 0.007). Italic fond represents nominal significant association (two-sided P < 0.05).

When we reproduced the organ-age estimation process in the FHS, the number of proteins selected in at least 90% ( ≥ 450) of the 500 bootstrapped LASSO iterations in the FHS training set was much smaller than in the UKB training set, even though the platform covered more than 5,000 proteins (Supplementary Data 10). The correlations of plasma-based brain age and chronological age were 0.79 and 0.76 in the training and testing set respectively, were a bit of lower than the correlations between organismal or conventional age and chronological age (Pearson r = 0.87-0.93). Unlike the FHS external validation results, the reproduced brain age was not significantly associated with verbal memory (effect size = −0.05, 95% CI = −0.09 to −0.01, P = 0.02) but associated with the language domain with multiple testing adjustment (effect size = −0.08, 95% CI = −0.12 to −0.04, P = 2.96 × 10^-4, Supplementary Data 11). The risk of having AD was higher as the plasma-based brain age increased by 1-SD (HR = 1.68, 95% CI: 1.44–1.96). Meanwhile, the risk of death was increasing along with the plasma-based brain, organismal, or conventional age increment (HR ranged from 1.46 to 1.61). No significant associations were found between the abovementioned three organ aging with diabetes, myocardial infarction, or stroke (Supplementary Data 12).

Function analysis of proteins comprising the brain aging predictor

We performed pathway enrichment analysis to investigate the potential functions of proteins contributing to the plasma-based brain age predictor. As shown in Supplementary Data 13, many of these proteins were involved in the synaptic pathways, including the KEGG synaptic vesicle cycle (multiple testing adjusted P = 2.63×10^-4) and Reactome pathways related to transmission across chemical synapses and the neuronal system (multiple testing adjusted P = 4.51 × 10⁻⁶ and 8.40 × 10⁻⁶, respectively). In contrast, pathways associated with organismal or conventional proteomic age were predominantly involved in cell surface receptor signaling, cytokine-cytokine receptor interaction, the immune system, and extracellular matrix organization.

Discussion

In this study, we developed a brain-specific aging proteomic clock and evaluated its associations with cognitive function and age-related chronic conditions in two prospective cohorts. Our findings indicate that accelerated brain aging is associated with lower cognitive performance and a higher risk of certain aging-related outcomes, including all-cause mortality and AD. Furthermore, our findings highlight the potential use of the proteomic brain clock as a valuable tool for monitoring overall health and identifying individuals at increased risk of age-related diseases.

Plasma proteins have emerged as valuable biomarkers for age-related diseases, particularly for disease risk stratifications, with several studies in the UK Biobank providing supporting evidence^4,12,14,16. Plasma-based brain age—defined as the gap between chronological age and predicted age based on brain-specific proteins using methods such as LASSO or XGBoost^14,16, has been associated with incident AD and mortality. Meanwhile, other researchers have proposed a proteomic clock, a set of proteins directly predicting mortality and age-related diseases using Cox models^4,12. In our study, we used a LASSO-derived plasma-based brain age measure as a predictor of both cognitive function and age-related disease outcomes. Because cognitive impairment often precedes clinical AD diagnosis by several years, our findings suggest that plasma-based brain age is associated with early signs of AD and may represent a potential therapeutic target for slowing disease progression. Another strength of our work is the use of the FHS as an external validation and replication cohort, which uses an Olink HT platform and on average 10 years older, thereby enhancing the robustness and generalizability of our results.

Compared to previous studies, we confirmed that plasma-based brain age is widespread, manifesting not only as advanced aging in patients with brain-related diseases but also in those with diseases affecting other systems, such as the heart⁶. Our finding that deaccelerated brain aging is associated with a lower likelihood of mortality further supports the idea that a youthful brain and immune system promote disease-free longevity¹⁴. Additionally, our findings suggest that cognitive function is significantly influenced by brain-specific aging proteins, as well as by a broader set of aging-related proteins. This suggests a more systemic approach to aging, where organismal aging proteins could also impact cognitive function, especially on the attention and concentration domain. This phenomenon may arise from protein-protein interactions or molecular communication between organs^42–44, highlighting the presence of shared proteomic biomarkers that reflect various aging phenotypes.

Traditional risk factors, such as diet and smoking, are linked to proteomic brain-specific aging. Smoking and alcohol consumption were associated with accelerated brain aging, while vigorous physical activity or a healthy diet (e.g., more vegetable consumption, less sugar-sweetened beverage consumption) was associated with a “slower” or delaying brain aging^14,15. Lifestyle interventions that address multiple modifiable traditional risk factors may benefit brain health, enhance cognitive performance, and ultimately help prevent age-related diseases^45,46.

Functional analysis of plasma-based brain age proteins highlights the involvement of synaptic vesicle cycling, transmission across chemical synapses, and neuronal system pathways in cognitive function and disease development. Because most brain-specific proteins (96 out of 102) were included in the enrichment analysis, the results did not pinpoint a small set of promising therapeutic targets or narrow the pathways relevant to slowing brain aging. To gain further insight, we examined the top three proteins contributing most strongly to the brain age model: myelin oligodendrocyte glycoprotein (MOG), neurofilament light chain (NEFL), and brevican (BCAN). MOG is implicated in myelin damage and subsequent neuronal and axonal degeneration⁴⁷, processes that underlie cognitive impairment in AD. NEFL has been associated with APOE4 genetic status and is a well-established marker of neurodegeneration^48–50. BCAN is linked to alterations of the extracellular matrix during aging and affects multiple cortical and subcortical structures^51,52, which may also contribute to cognitive decline.

Another potential mechanism by which brain-specific proteins influence cognitive function involves epigenetic changes. Previous studies have shown that some age-related DNA methylation sites are tissue-specific^53,54, suggesting a potential interplay between epigenetic and proteomic aging clocks in regulating brain health. In contrast, organismal proteins were enriched in Cytokine-cytokine receptor interaction, cell adhesion, cell surface receptor signaling pathways, indicating shared molecular mechanisms underlying aging in the brain and other organs. Additionally, a single-cell spatial transcriptomics atlas of the aging mouse brain across the lifespan has revealed complex cellular changes and cell-cell interactions associated with brain aging⁵⁵. Such studies may provide insights into how non-organ-specific aging proteins contribute to brain aging through intercellular interactions and broader systemic aging processes.

The clinical application of brain-specific aging measures holds promise for advancing research, as human plasma is easily accessible, and the cost of proteomic assays is expected to decrease with technological advancements. Additionally, the computational burden of brain age models is relatively low, as they are built using linear combinations. Implementing brain age assessments in clinical settings could help identify middle-aged individuals who are disease-free but at high risk of developing cognitive decline due to accelerated brain aging⁵⁶. Beyond serving as an aging biomarker, brain-specific proteomics also facilitates the identification of disease-associated proteins and their post-translational modifications, offering potential new druggable targets for therapeutic interventions^57,58. Early interventions or drug targets directed at brain-specific biomarkers can help slow down the aging process and prevent age-related diseases⁵⁹.

We acknowledge several limitations of the current study. First, we used GTEX mRNA expression data to define brain-specific proteins. The GTEX samples ranged in age from 20 to 70 years, with over two-thirds being older than 50 and a majority having chronic conditions such as heart or cerebrovascular diseases. As a result, relative gene expression levels across organs might have been influenced by aging and disease conditions. However, previous studies have demonstrated the feasibility and acceptability of GTEX-derived organ-specific definitions^15,17. Second, our aging models were developed using approximately 3,000 plasma proteins on the Olink platform, with validation conducted on ~2700 proteins measured using the Olink HT platform in the FHS. Given the smaller sample size of ~2000 participants, this validation cohort may have been slightly underpowered. The accuracy and robustness of these proteomic aging models could be improved with the inclusion of a broader range of proteins. Moreover, our brain aging model has not been compared to existing prediction models of blood-based biomarkers for neurodegenerative disorders. Third, although we adjusted sex in developing the aging model, the sex differences of brain aging were unexplored. Fourth, the association of brain aging and cognitive function was assessed cross-sectionally. Future longitudinal studies incorporating repeated cognitive assessments and proteomics are needed to better understand temporal relationships. Lastly, the study populations we used were predominantly of European ancestry. Therefore, further validation in diverse race/ethnicity populations is necessary to assess the generalizability of our findings.

Conclusions

In summary, we found that accelerated brain aging is associated with poorer cognitive performance cross-sectionally and a higher risk of age-related diseases across two longitudinal cohorts. These findings highlight the importance of understanding the molecular mechanisms driving brain aging and its role in cognitive decline and disease susceptibility. Our findings may help clinicians implement early interventions for individuals who are at risk of brain-related disorders.

Author contributions

B.W. and H.L. conceived the study idea. B.W. and H.D. conducted data analyses and wrote the first draft of the manuscript. D.Q., M.S.T., and J.M.M. reviewed the manuscript critically for important intellectual content. H.L. and J.M.M. supervised the work and made significant contributions to the interpretation of the results and editing of the manuscript. All authors read and approved the final version of the manuscript.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data that support the findings of this study are available from the UK Biobank and the Framingham Heart Study. The UK Biobank data generated and/or analyzed during current study are not publicly available for privacy reasons, but are available in the UK Biobank data repository with permission of the UK Biobank (https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access). We accessed the data under Application Number 76269. The source data for Fig. 2 is in Supplementary Data 14, for Fig. 3 is in Supplementary Data 9, for Fig. 4 is in Supplementary Data 15.

Code availability

The code used for data analysis is 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/s43856-025-01268-w.