Proteomic landscape of multidimensional aging phenotypes

Abstract

Background

Proteomic signatures of aging hold promise for advancing our understanding of aging evaluation and guiding targeted therapy. Despite this potential, the proteomic landscape of multidimensional aging phenotypes remains inadequately characterized. We aimed to identify the potential proteomic biomarkers of aging process and decipher their molecular mechanisms.

Methods

We analyzed 2920 plasma proteomic biomarkers from 48,728 participants in the UK Biobank. The multidimensional aging phenotypes included Klemera and Doubal’s method biological age (KDM-BA) acceleration, PhenoAge acceleration, frailty index, leukocyte telomere length (LTL), and healthspan. Two-sample Mendelian randomization (MR) analyses were performed to determine the causal effect of plasma proteome on the multidimensional aging phenotypes, and replicate the identified proteomic signatures in the FinnGen cohort. Multivariable linear regressions were used to explore the phenotypic associations between plasma proteome and multidimensional aging phenotypes. We then applied a series of bioinformatic approaches to elucidate the biological function and drug targets of the identified proteins. Multi-omics data were further leveraged to decipher the genetic mechanisms and metabolic pathways of aging process.

Results

We found that genetically determined levels of 17, 37, 12, 18, and 1 proteins were causally linked to KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan, respectively. Replication in the FinnGen cohort confirmed a subset of these associations. We observed significant phenotypic associations for 2,186, 2,152, 1,459, 668, and 545 proteins with KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan, respectively. Our integrative analysis identified 71 distinct plasma proteins associated with multidimensional aging phenotypes, of which 12 are promising candidates for drug targeting, primarily involved in inflammatory processes and cellular senescence. Moreover, we identified 22 genetic variants that may regulate these protein abundances in the context of aging, complemented by metabolomic profiling that highlights several metabolic pathways mediating the proteins and aging.

Conclusions

Our findings facilitate a more comprehensive understanding of the proteomic landscape of the multidimensional aging phenotypes, thereby providing an opportunity for personalized monitoring of aging and effective therapeutic strategies in aging-related diseases.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-025-01558-x.

Background

Aging is a process marked by the accumulation of senescent cells and a gradual decline in normal physiological functions, resulting in an increased risk of aging-related diseases and mortality [1]. Since aging is a complexly multisystemic process, a single biomarker may not fully and accurately represent the entire landscape of an individual’s aging experience, due to the inherent variability among cells, tissues, and organs [2]. Quantifying the aging process is not straightforward. A variety of domain-specific aging metrics have been developed and widely employed, including chronological measures like the length of time from birth until the onset of a major disease (healthspan) [3], indicators of cellular deterioration such as leukocyte telomere length (LTL) [4], biological age assessments like PhenoAge acceleration [5] and Klemera and Doubal’s method biological age (KDM-BA) acceleration [6], and comprehensive traits such as the frailty phenotype, which includes multiple indicators of functional impairment [7]. These metrics collectively capture multiple dimensions of aging, ranging from functional and phenotypic aspects to molecular perspectives. Thus, identifying central pathogenic factors and biological mechanisms of multidimensional aging phenotypes can inform preventive programs against aging and extend health expectancy.

Blood proteins, which enter circulation through active secretion or cellular leakage, provide a comprehensive overview of health and diseases states [8] and serve as a significant reservoir of biomarkers and therapeutic targets, offering greater accessibility and the highest predictive power for various diseases [9]. Plasma protein levels can offer a more direct understanding of the mechanisms and functions related to aging biology [10], and loss of proteostasis is recently raised as a primary hallmark of aging [11]. Using blood proteome to predict chronological age or mortality, several prior studies have identified aging-related proteomic biomarkers and utilized them to create proteomic age clocks for predicting the risk of certain diseases and mortality [12–16] (Wang S, Rao Z, Blaes AH, Coresh J, Joshu CE, Pankow JS, Thyagarajan B, Ganz P, Guan W, Platz EA, et al. Proteomic aging clocks and the risk of mortality among longer-term cancer survivors in the atherosclerosis risk in communities (ARIC) study. medRxiv. 2024 unpublished). There is also evidence depicting the proteomic signatures of healthy longevity [17].

Recent advances in large-scale proteomics have further enhanced our understanding of aging biology by associating specific protein signatures with particular aging phenotypes. For example, a study used a panel of 2920 proteins in the UK Biobank cohort and identified distinct proteomic signatures linked to frailty and its progression to adverse outcomes, highlighting proteins such as GDF15 and CD300E [18]. In parallel, another investigation employed proteome-wide MR to pinpoint 22 plasma proteins causally associated with telomere length, prioritizing markers like APOE and SPRED2 [19]. Similarly, research focused on healthspan developed a proteomic-based healthspan score that robustly predicted mortality and various age-related conditions [20]. Broader studies of the plasma proteome across the lifespan have revealed non-linear changes in protein expression across different life stages, and further work has demonstrated that proteomic biomarkers can reliably predict overall health- and life span [21, 22]. Despite these significant insights, the complete proteomic landscape that characterizes multidimensional aging metrics remains elusive. Moreover, while previous studies have highlighted associations between specific proteins and individual aging phenotypes, they have generally focused on isolated dimensions of aging. It is therefore imperative to elucidate both the upstream regulators and downstream consequences of these proteomic biomarkers to inform evaluation, intervention, and treatment strategies for aging-related diseases.

To address these gaps in knowledge, using plasma proteome data from the UK Biobank that involved large-scale measurements of approximately 3000 proteins in 50,000 participants, we aimed to (1) use two-sample MR analyses to determine the causal effects of plasma proteome from the UK Biobank Pharma Proteomics Project (UKB-PPP) cohort on multidimensional aging phenotypes, including KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan, and replicate the identified proteomic signatures in the FinnGen cohort; (2) investigate the phenotypic associations between plasma proteome and the multidimensional aging phenotypes; (3) elucidate the biological function and enrichment analyses of the potential proteomic biomarkers; (4) systematically assess potential adverse effects of the identified proteins on a wide range of disease; and (5) decipher the genetic mechanisms and metabolic pathways of the potential proteomic signatures of aging. A summary of the framework is illustrated in Fig. 1.

Fig. 1.

Workflow of the study. Data utilized in the present study, including information on multidimensional aging phenotypes and plasma proteome from the UK Biobank and FinnGen cohort. Step 1: we investigated the genetic and phenotypic effects between plasma proteome and five multidimensional aging phenotypes, including biological age metrics (KDM-BA acceleration, and PhenoAge acceleration), frailty index, LTL, and healthspan, and further replicated the identified proteomic signatures in the FinnGen cohort. Step 2: we elucidated the biological function and enrichment analyses of the potential proteomic biomarkers. Step 3: we systematically assessed potential adverse effects of the identified proteins on a wide range of disease, and performed drug repositioning analysis using data from OpenTargets. Step 4: we deciphered the genetic mechanisms and metabolic pathways of the potential proteomic signatures of aging

Methods

UK Biobank population sample

The UK Biobank is a prospective cohort study that has collected extensive genetic and phenotypic data from 502,364 participants residing in the UK, all of whom were recruited between 2006 and 2010 [23]. The complete UK Biobank protocol can be accessed online (https://www.ukbiobank.ac.uk/media/gnkeyh2q/study-rationale.pdf). Our sample was limited to participants who had Olink Explore data and were randomly selected from the main UK Biobank population, totaling 48,728 participants.

Proteomic and metabolomic profiling

Plasma samples collected from 2006 to 2010 were stored at the UK Biobank at − 80 ℃ and in liquid nitrogen (LN₂). Blood samples from randomly selected participants were then transported to the Olink Analysis Service in Sweden. Proteomic profiling of plasma samples was conducted using the Olink Explore 3072 platform, which integrates four Olink panels (cardiometabolic, inflammation, neurology, and oncology) from 54,306 UK Biobank participants between April 2021 and February 2022, capturing a total of 2923 unique proteins. Protein data are presented as Normalized Protein eXpression (NPX) values on a log₂ scale, with additional information on sample selection, processing, and quality control available elsewhere [24]. The UK Biobank implemented rigorous quality control measures and external validation. After excluding proteins with missing data exceeding 30% (GLIPR1, NPM1, and PCOLCE), a total of 2920 unique proteins were included. Participants with missing protein data exceeding 50% were excluded, resulting a total of 48,728 participants. The remaining missing protein values were imputed using the k-nearest neighbors imputation function. Subsequently, NPX values were inverse rank normalized prior to downstream analyses.

Metabolomic profiling was performed using a targeted high-throughput nuclear magnetic resonance (NMR) metabolomics platform [25]. This platform enables the simultaneous quantification of 249 metabolic measures, which includes the concentrations of 165 metabolic compounds and 84 derived ratios. It covers routine lipids, lipoprotein subclass profiling (including lipid composition across 14 subclasses), fatty acid composition, and a variety of low-molecular weight metabolites such as amino acids, ketone bodies, and glycolytic metabolites. Plasma samples were collected from a randomly selected subset of approximately 280,000 participants. Before analysis, all metabolic traits were inverse rank normalized. Detailed protocols for sample collection and metabolomic quantification have been outlined in previous studies [26].

Multidimensional aging phenotypes

We assessed aging using multidimensional metrics including KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan. We employed a frailty index on the basis of the accumulation of deficits model [7], which has been previously validated in the UK Biobank [27]. The frailty index was calculated using 49 self-reported baseline data variables in UK Biobank. These variables encompassed various physiological and mental health domains, including symptoms, disabilities, and diagnosed diseases reported by the participants (Table S1). The frailty index was created using a complete-case sample that included data on all 49 individual components, expressed as a proportion of the total sum of deficits.

LTL was assessed using multiplex quantitative polymerase chain reaction (qPCR), represented as the T/S ratio (telomeric repeats to single-copy genes), which compares the telomere amplification product (T) to a single-copy gene (S). Further details on LTL measurements, including quality checks and adjustments for technical factors in UK Biobank participants, have been provided previously [4].

Consistent with previous studies [3, 28], we defined each individual’s healthspan as the age at which they first experienced any of the following conditions (collectively referred to as chronic diseases): cancer, diabetes, coronary heart failure, myocardial infarction, stroke, chronic obstructive pulmonary disease, dementia, and death. Specifically, if the participant had experienced onset of any abovementioned chronic diseases prior to or at baseline, the age at first diagnosis was used. If no disease was present, the baseline age at assessment was taken as the healthspan endpoint. Disease diagnoses were confirmed through inpatient hospital admission data and self-reported information collected via verbal interviews.

Biological age acceleration was calculated by taking the residuals of KDM-BA or PhenoAge after removing the influence of chronological age through a linear regression model, referred to as KDM-BA acceleration or PhenoAge acceleration [5]. KDM-BA and PhenoAge, derived from multisystemic chemical biomarkers, has become notably comprehensive metrics for evaluating biological aging [5, 29]. KDM-BA was developed from nine clinical biomarkers, including forced expiratory volume in the first 1.0 s (FEV₁), systolic blood pressure, total cholesterol, glycohemoglobin (HbA1c), albumin, creatinine, C-reactive protein, alkaline phosphatase, and blood urea nitrogen. PhenoAge was developed from nine blood chemistries, including albumin, alkaline phosphatase, creatinine, C-reactive protein, glucose, mean cell volume, red cell distribution width, white blood cell count, and lymphocyte proportion. KDM-BA acceleration and PhenoAge acceleration have been widely utilized and shown to effectively capture morbidity and mortality risks across diverse subpopulations in various countries [30, 31]. To address non-normality, all 5 multidimensional aging phenotypes were inverse rank normalized for analysis.

Genetic data sources

GWAS summary-level data for aging

We conducted genome-wide association studies (GWAS) by ourself for the multidimensional aging phenotypes, including KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan from the UK Biobank samples without proteomics data (n = 350,313, 311,471, 367,696, 355,755, and 367,873, respectively). GWAS was performed using REGENIE [32], controlling for age, age², sex, age × sex, twenty ancestral principal components, and genotyping array.

GWAS summary-level data for plasma proteins

In the discovery phase, we used GWAS summary statistics for plasma protein abundance derived from the UKB-PPP of 54,219 UK Biobank participants [24]. To validate our results, we extracted summary-level genetic association data on 2925 circulating proteins measured by Olink technique from an extensive pQTL study involving 619 Finnish participants, known as the FinnGen cohort [12].

GWAS summary-level data for parental lifespan

To provide independent validation of our findings, we utilized GWAS summary statistics for parental lifespan from Timmers et al. [33], which included approximately 1 million participants of European ancestry. This dataset represents a combined phenotype of mother’s and father’s lifespan, providing a complementary measure of human longevity that is genetically correlated with healthspan and other aging phenotypes.

Statistical analysis

Healthspan assessment and temporal considerations

In our primary analysis, we used baseline healthspan values for all participants. This cross-sectional approach was chosen to match our proteomic profiling, which was performed at a single time point during the baseline assessment. To address potential limitations of this cross-sectional design, we conducted sensitivity analyses incorporating longitudinal follow-up data. We performed additional GWAS using updated healthspan values that included incident disease diagnoses during UK Biobank follow-up. We then repeated our MR analyses using these updated GWAS summary statistics to validate our primary findings.

MR analyses

It is possible that the plasma proteins causally related to the multidimensional aging phenotypes might not be among those examined by the observational studies due to residual confounders. Therefore, we used the two-sample MR design for the primary analyses (plasma proteins versus the five aging phenotypes). We selected conditionally independent cis-protein quantitative trait loci (pQTL) from GWAS summary data of plasma proteins by using the clump function in “TwoSampleMR” R package. We set a linkage disequilibrium (LD) pruning r² threshold of 0.001, a window size of 1 Mb, and a p-value threshold of 5 × 10⁻⁸. Due to the limited sample size in the FinnGen pQTL analysis, we used a less stringent p-value threshold of 1 × 10⁻⁶. The instruments were selected to satisfy the three core MR assumptions: single nucleotide polymorphisms (SNPs) are strongly associated with protein but not with aging phenotypes and not with confounding factors. SNPs with indirect LD effects (r² > 0.8) were also removed if they were associated with the aging phenotypes. Furthermore, F-statistics was calculated to measure the power of instrumental variables (IVs), with a threshold below 10 suggesting greater bias [34]. We performed two-sample MR analyses to confirm the causal relationships between plasma proteins and the multidimensional aging phenotypes. The Wald ratio was used if only one pQTL was available for a given protein. When two or more genetic instruments were available, inverse-variance weighted (IVW) approach was applied. To correct for multiple comparisons across the proteome, we corrected the p-values using both the Bonferroni method and false discovery rate (FDR) method, with Bonferroni-corrected p < 0.05 or FDR q < 0.05 considering statistically significant. We used MR Steiger filtering to check whether the MR analysis estimates assessed the true causal direction [35]. To improve the accuracy and robustness of the genetic instruments, we used Cochran’s Q test for IVW model to assess heterogeneity [36]. MR-Egger regression was applied to detect the potential bias of directional pleiotropy [37]. The intercept in the Egger regression represents the average pleiotropic effect of all genetic variants. A value that significantly deviates from zero (p < 0.05) is taken as evidence of pleiotropy.

For the proteins of interest, we further performed sensitivity analyses to complement and enhance the reliability of the results. First, MR robust adjusted profile score (MR-RAPS) addresses both systematic and idiosyncratic pleiotropy, allowing for more robust inferences in MR analyses, especially when dealing with numerous weak instruments [38]. Second, the MR-Egger method estimates the causal effect using the slope coefficient from the Egger regression, offering a more robust estimate even when none of the IVs is valid [39]. Third, we conducted MR pleiotropy residual sum and outlier (MR-PRESSO) test to examine the presence of horizontal pleiotropy (p < 0.05) [40].

To validate the robustness of our MR findings, we performed additional two-sample MR analyses using parental lifespan GWAS summary statistics as the outcome. We applied the same analytical framework as described above, using cis-pQTLs from UKB-PPP as IVs. Proteins showing nominal significance (p < 0.05) in both the primary aging phenotype analyses and parental lifespan validation were considered as having robust evidence for association with aging.

Observational analyses

We presented descriptive statistics for categorical variables as counts with percentages, and for continuous variables, we reported the mean ± SD. Multivariable linear regression was performed to estimate the coefficients of plasma protein levels with five multidimensional aging phenotypes and chronological age. Each plasma protein was used as the independent variable and aging phenotypes as the response variable. The adjusted covariates included age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and body mass index (BMI). For proteomics-associations of chronological age, we adjusted for sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI. Next, we employed a linear mixed model explore the robustness of proteomics-aging associations, incorporating assessment center as a random effect. The Bonferroni method was used for multiple testing correction, with Bonferroni-corrected p < 0.05 as the significance level.

Phenome-wide association study

To explore potential adverse effects associated with promising proteins, we examined associations with the full range of clinical phenotypes encoded in the UK Biobank. Events were identified from inpatient hospital record International Classification of Diseases 10th Revision (ICD-10) codes, which were categorized into phenotypes (PheCODEs) using the PheCODE Map v1.2b1 [41]. Individuals were designated as cases if they had more than one ICD-10 code mapped to the corresponding PheCODE. The PheCODE Map includes exclusion criteria for each phenotype, identifying similar conditions that may suggest the likelihood of undiagnosed patients for that phenotype. Cox proportional hazards model was performed for each PheCODE, and protein level was treated as an independent variable and age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI as covariates. The Bonferroni method was used for multiple testing correction, with Bonferroni-corrected p < 0.05 as the significance level.

Linkage disequilibrium score regression

To estimate the shared heritability between the five multidimensional aging phenotypes, we applied cross-trait linkage disequilibrium score regression (LDSC) approach [42] based on GWAS summary statistics to calculate the genetic correlation coefficient (r_g). The method leverages the relationship between χ² statistics from GWAS and LD scores, where SNPs with higher LD scores (aggregated across neighboring variants within 1 centiMorgan windows) are expected to have larger χ² statistics under a polygenic architecture. The method regresses GWAS χ² statistics against LD scores (precomputed from UK Biobank in-sample LD panel) to quantify polygenicity and estimate heritability h²_SNP. The genetic correlation estimates range from − 1 to 1. Analyses were restricted to HapMap3 SNPs, and the major histocompatibility complex (MHC) region was excluded due to complex LD patterns.

Colocalization analysis

We used the COLOC method [43] to assess the probability of the same single-nucleotide variation being responsible for both changing aging and modulating the protein levels of a gene. We assigned the default prior probabilities to an SNP interrelated with an aging phenotype (p₁ = 1 × 10⁻⁴), an SNP with a significant pQTL (p₂ = 1 × 10⁻⁴), and an SNP associated with both traits (p₁₂ = 1 × 10⁻⁵). COLOC employs computed approximation Bayes factors and summary association data to generate posterior probabilities for the following five hypotheses: H0, no association with either GWAS or pQTL; H1, association with GWAS, not with pQTL; H2, association with pQTL, not with GWAS; H3, association with GWAS and pQTL, two independent SNPs; and H4, association with GWAS and pQTL, one shared SNP. The posterior probabilities (PP), denoted as PP0, PP1, PP2, PP3, and PP4, quantify the support for each hypothesis. A posterior probability of association of ≥ 0.80 for the last hypothesis (H4) is considered to provide strong evidence of colocalization.

Summary-data-based Mendelian randomization

Summary-data-based Mendelian randomization (SMR) [44] was used to test whether causal variants were associated with aging via their cis-regulated protein abundance. We used the pQTL results and the GWAS summary statistics of the five multidimensional aging phenotypes to perform SMR, with the default parameters recommended by the developers of SMR software v1.03 [44]. Only protein expression probes with at least one cis-pQTL at p < 5 × 10⁻⁸ were included. The heterogeneity in dependent instruments (HEIDI) test was used to test whether observed association between protein abundance and aging was likely due to a pleiotropic scenario, which was indicated by a HEIDI test (p < 0.05). We applied a conservative unadjusted threshold of p < 0.05 from HEIDI to indicate that the presence of linkage is likely to influence the primary SMR findings.

Plasma metabolomics analysis

To understand individual metabolite associations with aging, we first performed multivariate analyses of 249 NMR metabolic measures and the five aging phenotypes. Metabolome wide association study (MWAS) of the 5 multidimensional aging phenotypes was assessed using multivariable linear regression after adjustment for age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI. Bonferroni correction was used when accounting for multiple testing, with Bonferroni-corrected p < 0.05 denoting significance.

Mediation analysis

Casual mediation analysis was applied to investigate whether the associations between the identified proteins and aging were mediated by plasma metabolites in the UK Biobank, using “mediation” R package. The mediator model was fitted using linear regression to detect the associations between proteins and each metabolite with adjustment for age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI. The outcome model was fitted using linear regression to detect the associations between proteins and aging with adjustment for the same covariates set.

Pathway enrichment

We used the g:Profiler tool [45] to perform the functional enrichment and pathway characterization of the 71 distinct proteins that were identified in the MR analysis, which leverages on the diverse sources of biological evidence including Gene Ontology (GO) biological processes, cellular components, and molecular functions. In g:Profiler, multiple-testing correction was conducted by the FDR method at 5% threshold. MR-identified proteins after Bonferroni correction were used as the test set, with all proteins measured in the UK Biobank (n = 2920) serving as the background set.

The analysis of protein–protein interaction networks

To test the interactions of the identified proteins, the protein–protein interaction (PPI) network was constructed by querying the STRING database [46] (v.12.0; Homo sapiens, taxonomy ID: 9606) for proteins significantly relating to aging using a medium-to-high confidence interaction score threshold (combined score > 400). All default STRING evidence types were incorporated in the network construction, including experimental validation, database annotation, text mining, co-expression, neighborhood, gene fusion, and co-occurrence evidence. Proteins that cannot form network were removed. PPI network was visualized using igraph R package.

Tissue and cell type-specific enrichment analysis

Considering that different genes exhibit differential expressions across various tissues and cells, we performed tissue and cell type-specific enrichment analysis of the proteins identified by MR analysis. Tissue-specific expression enrichment analysis was carried out using gene expression data from the Genotype-Tissue Expression (GTEx) project [47]. The FUMA platform [48] utilized gene expression data from 30 main tissue types available in the GTEx v8 database. For detailed information, please refer to the original publication [47]. Genes with p-value less than 0.05 in each tissue after Bonferroni correction were defined as “tissue-specific.”

Cell type-specific expression enrichment analysis was performed using single-cell RNA sequencing data for the human liver shared by Martin et al. [49]. This study assessed the cell type-specific expression of target genes potentially causally linked to human liver transcriptional levels. The single-cell RNA sequencing data for liver tissues included 167,598 cells. Initially, the raw single-cell RNA sequencing data underwent data preprocessing and transformation using R package Seurat v5.0.2. Genes with counts less than 3 in a single cell and cells with fewer than 50 unique feature counts were removed. To examine whether specific liver cell types exhibited higher expression of identified causal protein-coding genes, differential expression analysis was conducted using the Wilcoxon rank-sum test, comparing the gene expression levels between one cell type and the other cell types. Genes with an average log₂ fold change (log₂FC) greater than 0.5 and Bonferroni-corrected p-value less than 0.05 were identified as enriched genes for that specific cell type.

Protein druggability

Proteins encoded by genes identified in the MR analysis were annotated with drug tractability information based on information provided by OpenTargets [50]. OpenTargets tractability system stratified drug targets into nine mutually exclusive groups (termed “buckets”) based on the drug type and the stage of the drug discovery pipeline.

Results

Study participants

This study included 48,728 participants (54.0% women; mean (SD) age, 56.8 (8.2) years) with proteomic data from the UK Biobank. Population characteristics of the study are presented in Table S2. The genetic correlations between the five multidimensional aging phenotypes in the UK Biobank were generally modest (absolute r_g ranged from 0.10 to 0.62) (Table S3). Age distribution across five multidimensional aging phenotypes is illustrated in Fig. S1.

Causal effect of plasma proteome on aging

We investigated potential causal relationships linking proteomic biomarkers to the multidimensional aging phenotypes. Using the GWAS summary-level data of plasma proteome from the UKB-PPP cohort, we selected conditionally independent cis-SNPs, defined as any variant located within a ± 1 Mb region of the protein-coding gene, that was related to plasma levels (p < 5 × 10⁻⁸). To make sure that the samples of the exposures were independent from those of the outcomes, we conducted GWAS of the multidimensional aging phenotypes from the UK Biobank non-overlapping samples (Fig. 2). For KDM-BA acceleration, the heritability h²_SNP estimated by LDSC was 0.096 (standard error (SE) = 0.005), with mean χ² and genomic control inflation factor (λ_GC) estimated at 1.858 and 1.576, respectively, and the LD score intercept of 1.120 (SE = 0.016). For PhenoAge acceleration, the h²_SNP was 0.158 (SE = 0.009), mean χ² and λ_GC at 2.165 and 1.643, respectively, and an intercept of 1.146 (SE = 0.027). For frailty index, the h²_SNP was 0.122 (SE = 0.004), mean χ² and λ_GC at 1.993 and 1.768, respectively, and an intercept of 1.058 (SE = 0.015). For LTL, the h²_SNP was 0.073 (SE = 0.008), mean χ² and λ_GC at 1.576 and 1.222, respectively, and an intercept of 0.997 (SE = 0.022). For healthspan, the h²_SNP was 0.007 (SE = 0.001), mean χ² and λ_GC at 1.068 and 1.065, respectively, and an intercept of 1.011 (SE = 0.008).

Fig. 2.

Genome-wide associations of multidimensional aging phenotypes. Circular Manhattan plot depicting the genomic landscape of the multidimensional aging phenotypes across autosomal chromosomes (Chr1–Chr22). The outer ring displays chromosomal positions with SNP density indicated by color intensity. The inner concentric tracks show genetic associations with five multidimensional aging phenotypes. Red dots throughout the inner rings represent significant genetic associations with these aging phenotypes (p < 5 × 10⁻⁸)

We next applied two-sample IVW or Wald-scores MR analyses to investigate potential causal effects on aging. We found 17 plasma proteins for KDM-BA acceleration, 37 for PhenoAge acceleration, 12 for frailty index, 18 for LTL, and 1 for healthspan surpassed the statistical threshold for significance (Bonferroni-corrected p < 0.05, Fig. 3). Using a less stringent threshold for significance (FDR q < 0.05), we found 50 plasma proteins for KDM-BA acceleration, 115 for PhenoAge acceleration, 26 for frailty index, 32 for LTL, and 1 for healthspan. Taken together, we identified 71 distinct proteins of aging following Bonferroni correction, and 180 distinct proteins of aging following FDR correction. The full summary of MR results is shown in Table S4. We applied MR Steiger filtering to evaluate whether the MR estimates reflected the true causal direction, and our analyses revealed no evidence of reverse causality. The F-statistics was used to evaluate the instrument strength. For the 224 MR pairs with at least one IV, the F-statistic values were all ≥ 30, indicating that these SNPs were suitable IVs. The heritabilities estimated from LDSC for the 71 distinct proteins of aging are displayed in Table S5.

Fig. 3.

Causal effects of plasma proteins on multidimensional aging phenotypes. a–e Volcano plots illustrating the MR analysis on the associations between plasma proteins of the UK Biobank cohort and the five multidimensional aging phenotypes. f Venn diagram visualizing the overlap of MR-identified proteins (Bonferroni-corrected p < 0.05) across five multidimensional aging phenotypes. g Forest plot showing the causal effects of 71 distinct proteins on aging phenotypes identified through MR. Effect sizes (β) and 95% CIs are shown for each protein-aging association. Dot size represents the number of instrumental variables (IVs) used

We conducted several sensitivity analyses to confirm our putative causalities obtained by the genetic association. First, heterogeneity in the IVW estimates was detected by Cochran’s Q test; we only observed 32 out of 224 MR pairs suggesting moderate evidence of heterogeneity in the SNP effect estimates (p < 0.05), indicating most causal proteins exhibited well validity (Table S4). Second, MR-Egger intercepts showed 106 out of 112 MR pairs closing to zero, indicating that no significant pleiotropy was examined (Table S4). Third, the global MR-PRESSO test [38] did not find any evidence of horizontal pleiotropy for almost all potential proteins (Table S6). Fourth, the MR-RAPS was further performed to model the distribution of the estimates from invalid IVs, where the causal effects were consistent with estimates of IVW (Table S6). Overall, the sensitivity analyses validated the reliability of the potential causal effects found in our MR results.

Validation of the identified plasma proteins

In the validation phase, we sought to replicate the identified plasma proteins that reached Bonferroni significance threshold in the MR analysis in the FinnGen cohort. We found 4 plasma proteins for KDM-BA acceleration, 5 for PhenoAge acceleration, 1 for frailty index, 3 for LTL, and 0 for healthspan surpassed the statistical threshold for significance (uncorrected p < 0.05, Table S7). Taken together, we totally identified 71 distinct plasma proteins associated with aging, of which 12 proteins were strongly validated by different study design. These findings underscore the substantial relationships between the identified proteins and aging.

Validation using longitudinal healthspan data

To validate our healthspan findings, we performed sensitivity analyses using longitudinal follow-up data. The longitudinal healthspan GWAS incorporating 12.5 years of follow-up (n = 367,873) showed high genetic correlation with our baseline healthspan GWAS (r_g = 0.83). In MR analyses using the longitudinal healthspan GWAS, HLA-DRA remained the only protein significantly associated with healthspan at the Bonferroni-corrected or FDR-corrected threshold (β = − 0.02, p = 2.33 × 10⁻¹⁴, number of IVs = 21), consistent with our primary analysis (β = − 0.02, p = 3.26 × 10⁻⁷, number of IVs = 21).

Independent validation using parental lifespan

To further validate our findings using an independent aging-related phenotype, we performed MR analyses with parental lifespan GWAS data (n ≈ 1,000,000). Among the 71 distinct proteins identified in our primary analyses, 12 proteins (16.9%) showed consistent directional effects and achieved nominal significance (p < 0.05) with parental lifespan. These robustly validated proteins included APOE (β = − 0.122, p = 4.24 × 10⁻⁵), FURIN (β = 0.110, p = 3.16 × 10⁻⁷), SDCCAG8 (β = − 0.075, p = 8.19 × 10⁻⁵), CELSR2 (β = − 0.052, p = 0.007), HLA-E (β = 0.035, p = 0.009), ZPR1 (β = 0.199, p = 0.009), CEP170 (β = 0.149, p = 0.011), CD46 (β = − 0.111, p = 0.011), ASGR1 (β = 0.062, p = 0.02), UMOD (β = 0.014, p = 0.034), CXCL8 (β = 0.106, p = 0.041), and ATXN2L (β = 0.164, p = 0.048) (Table S8).

Phenotypic association between plasma proteome and aging

In the discovery phase, the proteomic signatures of the five multidimensional aging phenotypes were characterized through multivariable linear regression analyses. Of the 2920 proteomic biomarkers tested by Olink technique in the UK Biobank, after adjusting for age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI, we identified 2186, 2152, 1459, 668, and 545 proteins with KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan, at a Bonferroni-corrected p < 0.05, respectively (Fig. S2a–e and Table S9). For comparison, we identified 1801 proteins significantly associated with chronological age after Bonferroni correction (Fig. S2f and Table S9). The number of significant proteins overlapping with chronological age and KDM-BA acceleration, PhenoAge acceleration, frailty index, LTL, and healthspan were 1513, 1440, 1119, 515, and 515, respectively. Most proteins showed negative correlations with healthspan (n = 462), with GDF15 showing the strongest effect, followed by REN and LGALS4. Other proteins (n = 83) had positive correlations, such as PLA2G7 and FGFBP1. Most proteins showed positive correlations with frailty index (n = 1318), with GDF15 again showing the strongest effect, followed by LEP and LGALS4, while other proteins (n = 141) had negative correlations, such as CA14 and ITGA11. Comparisons of z-scores between the proteins associated with the aging phenotypes and the chronological age are displayed in Fig. S3, showing that proteomic signatures of KDM-BA acceleration, PhenoAge acceleration, and frailty index were positively associated with that of chronological age (r = 0.55, 0.49, and 0.53, respectively), while proteomic signatures of healthspan and LTL were negatively associated with that of chronological age (r = − 0.54 and − 0.29, respectively). We have conducted additional analyses using linear mixed models that account for potential confounding factors. These results remained largely consistent with our original findings, supporting the robustness of the observed associations (Fig. S4).

Subgroup analyses stratified by sex (female vs. male) and age (< 65 vs. ≥ 65 years) were performed for each protein exhibiting a significant association with five multidimensional aging phenotypes. Following Bonferroni correction, we identified 714 proteins with significant sex-interaction effects. Among these significant proteins, 83.2% (594) of proteins shared between sexes, with 10.6% (76) and 6.2% (44) showing male-specific and female-specific associations, respectively (Fig. S5). For instance, scavenger receptor class F member 2 (SCARF2) demonstrated a significant effect modification with PhenoAge acceleration (β (95% CI) for female = 0.07 (0.06 to 0.09); β (95% CI) for male = 0.17 (0.15 to 0.19); p-value for interaction = 1.23 × 10⁻⁶⁸) (Table S10). Similarly, stratification by age identified 790 significant interactions (Fig. S6). For instance, growth/differentiation factor 15 (GDF15) demonstrated a significant effect modification with PhenoAge acceleration (β (95% CI) for < 65 years = 0.25 (0.24 to 0.26); β (95% CI) for ≥ 65 years = 0.45 (0.42 to 0.47); p for interaction = 8.47 × 10⁻⁵¹). Specifically, among the 1459 proteins significantly associated with frailty index, 33 proteins (2.3%) exhibited significant age-specific associations (Bonferroni-corrected p for interaction < 0.05) (Table S11).

Enrichment analyses and biological functions

We elucidated the functional enrichment and pathway characterization of the MR-identified 71 distinct proteins using g:Profiler tool [45]. After restricting the analysis to GO pathways, we captured 26 enriched pathways (FDR q < 0.05). Biological pathways included regulation of immune response (FDR q = 0.04), and regulation of cell adhesion (FDR q = 0.04, details in Fig. 4a and Table S12). Notably, cellular components included MHC protein complex (FDR q = 0.007; HLA-A, HLA-E, and HLA-DRA). These findings highlight that inflammatory response may play a crucial role in the aging.

Fig. 4.

The biological function of the potential plasma proteins of aging. a Enrichment of protein-coding genes associated with the five multidimensional aging phenotypes in g:Profiler. Gene sets and pathways results were corrected for multiple testing by the FDR method at 5% threshold. GO:BP, Gene Ontology Biological Process; GO:CC, Gene Ontology Cellular Component; GO:MF, Gene Ontology Molecular Function. b Protein-protein interaction (PPI) network between the 71 distinct proteins of the five multidimensional aging phenotypes. The nodes represent proteins, with width of the edges connecting the nodes indicating a greater degree of connectivity. c Tissue-specific expression of protein-coding genes associated with the five multidimensional aging phenotypes. The left section displays a hierarchical clustering heatmap of gene expression across 30 different human tissues (rows) derived from GTEx v8 data. The dendrogram at the top groups genes with similar cross-tissue expression patterns. The right panel displays statistical significance of tissue-specific expression, with the dashed vertical line indicating the Bonferroni-adjusted significance threshold (p < 0.05). d Uniform manifold approximation and projection visualization of single-cell RNA sequencing data for human liver. The color of each dot indicates the cell type. e Differential expression of 66 protein-coding genes showed enrichment in the relevant cell types. The y-axis displays the average log₂ fold change (log₂FC), with positive values (above zero) indicating upregulation and negative values indicating downregulation of gene expression compared to other cell types

To investigate the potential interactions between the identified proteins, a PPI network was developed using the STRING database [46]. We observed the 87 interactions between the potential proteins (Fig. 4b, Table S13). Individual proteins with the greatest numbers of connections to other proteins were NFKB1 (nuclear factor kappa B subunit 1; a key transcription factor that functions as a master regulator of inflammation and immune responses [51]), CXCL8 (C-X-C motif chemokine ligand 8; also known as interleukin-8, a pro-inflammatory chemokine [52]), and APOE (apolipoprotein E; a major lipid transport protein and involved in immune response [53]). The three proteins emphasized the significance of inflammation and cellular senescence in human aging.

To detect if the identified proteins of aging were enriched in specific human tissues, we conducted tissue enrichment analysis by using the FUMA platform that included gene expression data from 30 main tissues (GTEx v8) [48]. In the tissue enrichment analysis, we used the set of proteins identified at an FDR-corrected significance threshold (FDR q < 0.05) from the MR analyses. Differential expression analyses revealed that protein-coding genes exhibited differential expression only in liver tissues (Bonferroni-corrected p < 0.05) (Fig. 4c, Table S14–15). Moreover, we conducted single-cell type expression analysis using single-cell RNA sequencing data of human liver [49]. Cells were clustered into cell types (Fig. 4d) and we found that 66 protein-encoding genes significantly exhibited cell type-specific enrichment in the human liver. For instance, APOE was primarily enriched in macrophages, and LEPR enriched in hepatocytes (Fig. 4e, Table S16).

Phenotype-wide association analyses and druggability evaluation

To obtain insight into health conditions associated with potential plasma proteins, we conducted a multivariable phenotype-wide association analyses (PheWAS) for 71 distinct proteins related to aging. A total of 1156 unique phenotypes defined into 16 disease classes were extracted from medical records and categorized according to the PheCODE Map [41]. Cox proportional hazard models were performed with each PheCODE diagnosis as an endpoint and each protein as the independent variable. Overall, a total of 2716 (3.3%) PheWAS pairs reached the Bonferroni-corrected significance threshold (Fig. 5a and Table S17), of which 2447 pairs exhibited deleterious and 269 beneficial effects. A total of 64 out of 71 proteins showed at least one significant association with PheCODE diagnosis after Bonferroni correction. Figure 5b–e provides a detailed view of the strongest incident disease associations for four exemplar biomarkers, with additional examples available in Table S17. More specifically, ASGR1 was associated with 202 health conditions such as chronic renal failure, type 2 diabetes, and hypertension. TNFRSF6B was related to 196 diseases spanning 15 PheCODE chapters, of which 196 conditions including chronic renal failure, type 2 diabetes, hypertension, and pneumonia being deleterious. There were 157 diseases related to HLA-E, including chronic renal failure, gout, and heart failure. HYAL1 was related to 74 diseases, such as chronic renal failure, type 2 diabetes, essential hypertension, and hypercholesterolemia.

Fig. 5.

The potential proteomic biomarkers of aging for future disease onset across a spectrum of diseases. a Total number of incident disease significantly associated with plasma proteomic biomarker. The disease outcomes were defined based on PheCODE, with a total of 1156 diseases tested for association. The color coding indicates the proportion of associations coming from each PheCODE chapter. b–e Phenome-wide association analyses for significant proteomic biomarkers: b ASGR1, c TNFRSF6B, d HLA-E, e HYAL1. All models used Cox proportional hazards model with adjustment for age, sex, ethnicity, Townsend deprivation index, education level, smoking status, alcohol consumption, healthy diet score, and BMI. The dotted line corresponds to Bonferroni-corrected p < 0.05

In druggability evaluation using the Open Targets Platform [50], we found that 12 out of 71 proteins have been targeted for drug development (Table S18). For instance, drugs named AZELIRAGON targeting AGER have been developed to treat Alzheimer’s disease and diabetic nephropathy. Drugs named EDASALONEXENT as an inhibitor targeting NFKB1 have been approved for the treatment of type 2 diabetes and Duchenne muscular dystrophy. Some drugs targeting PARP1 have been developed to treat several subtypes of cancer.

Genetic mechanism for the potential plasma proteins

To identify causal variants that underlie proteomic associations with aging, we employed two bioinformatic approaches to investigate whether risk variants influence aging etiology by affecting protein expression. First, we performed colocalization analyses to detect whether the MR-identified associations between the potential proteins and the five multidimensional aging phenotypes shared the same causal variants. We found 26 out of 85 protein-aging pairs showed evidence of colocalization, as determined by a posterior probability of association (PP.H4) ≥ 0.80 for at least one common genetic variant in the examined genomic region influencing both protein and aging phenotypes (Table S19). Colocalization results included UMOD, HEXIM1, ACP1, EFEMP1, ATXN2L, FURIN, CELSR2, APOE, NT5C3A, IL1RN, KIT, SERPINF2, CEP170, SERPINA1, ARFIP1, NFATC1, TRDMT1, PARP1, TK1, TYMP, RPA2, COMMD1, and HLA-DRA. For example, rs429358, a missense variant in APOE, which is a major apolipoprotein that mediates cholesterol and lipid transport in the bloodstream and central nervous system, plays a critical role in neurological health and is strongly associated with Alzheimer’s disease risk [54]. Furthermore, we used SMR and the HEIDI test [44] based on summary-level data from GWAS and cis-pQTL studies to test whether the causal variants were associated with aging via their cis-regulated protein abundance. We found 81 out of 85 protein pairs passed the SMR test (p < 0.05), of which 11 passed the HEIDI test (p > 0.05), suggesting genetically regulated protein abundance levels modulating aging. Taken together, we identified that 22 causal variants drive the specific protein level of aging (UMOD, HEXIM1, ACP1, EFEMP1, ATXN2L, FURIN, CELSR2, APOE, NT5C3A, IL1RN, KIT, SERPINF2, CEP170, SERPINA1, ARFIP1, NFATC1, TRDMT1, PARP1, TK1, TYMP, RPA2, and COMMD1).

Metabolites-mediated protein pathways

We used data from 274,355 participants of UK Biobank with 249 metabolic measures quantified by NMR, comprising 168 measures in absolute levels and 81 ratio measures. We systematically estimated the associations of the 249 NMR biomarkers with the five multidimensional aging phenotypes. Overall, we observed a total of 853 out of 1245 individual metabolite-aging associations passed Bonferroni-corrected significance threshold, of which 210 pairs for KDM-BA acceleration, 206 for PhenoAge acceleration, 202 for frailty index, 181 for healthspan, and 54 for LTL (Fig. 6a–b and Table S20).

Fig. 6.

The plasma metabolomic profiling of the multidimensional aging phenotypes and potential proteomic biomarkers of aging. a The number of significant metabolite associations for each aging phenotype from metabolite-wide association studies (MWASs), with colored bars representing different metabolite categories. b Manhattan plot for metabolite associations with each aging phenotype. c–g Chord diagrams showing the significant metabolite associations of the potential proteins for each aging phenotype. h–l Mediation analysis examining the relationship between plasma proteins, metabolites, and aging phenotypes. We selected several associations of plasma proteins, such as h ASGR1, i HLA-E, j HYAL1, k HLA-DRA, and l TNFRSF6B, with the multidimensional aging phenotypes that were mediated by the metabolites. Plasma metabolome consists of 249 metabolic traits (168 concentrations and 81 ratios). The legend identifies 47 metabolite categories. GlycA, glycoprotein acetyls; L_HDL_PL_pct, phospholipids to total lipids in large high-density lipoprotein percentage; L_LDL_C_pct, cholesterol to total lipids in large low-density lipoprotein percentage

Aging is commonly linked to metabolic and proteomic alterations. Therefore, we hypothesized that protein expression could influence the transport or synthesis of metabolites, affecting the concentration of circulating small molecules and ultimately leading to senescence. We applied two approaches relying on different assumptions to examine the mediating role of plasma metabolome on the proteomic signatures of aging. For the first approach, we used multivariable linear regression models to estimate the association between the 71 potential proteins and metabolites. Overall, we observed the 12,577 out of 17,679 protein-metabolite associations (Fig. 6c–g, Table S21), of which all metabolites were associated with 17 proteins that related to KDM-BA acceleration (Fig. 6c), 37 proteins that related to PhenoAge acceleration (Fig. 6d), 12 proteins that related to frailty index (Fig. 6e), 18 proteins that related to LTL (Fig. 6g), and 1 protein that related to healthspan was associated with 82 metabolites (Fig. 6f). For the second approach, we used mediation models to test whether the association between the identified proteins and aging phenotypes were mediated by the identified metabolic measures. Following Bonferroni correction, we found 9356 significant metabolic pathways linking 71 proteins and the five multidimensional aging phenotypes. Taking ASGR1as an example, the association with PhenoAge acceleration was mediated by creatinine, which accounted for 29% of the total effect on PhenoAge acceleration (Fig. 6h). Other proteins such as HLA-E, HYAL1, HLA-DRA, and TNFRSF6B are presented in Fig. 6i–l and Table S22.

Discussion

To the best of our knowledge, this represents the most extensive investigation so far that unveiled the proteomic landscape of healthy aging, denoted as functional, phenotypic, and molecular aging metrics. Using plasma proteome data on the UK Biobank, we identified 71 distinct plasma proteins genetically linked to multidimensional aging phenotypes and validated 12 of them in the FinnGen cohort. Furthermore, our validation using parental lifespan GWAS data from approximately 1 million participants provides independent confirmation of our key findings, with 12 proteins showing robust associations across different aging-related phenotypes. The identified proteins were involved in specific biological pathways related to inflammatory response and cellular senescence, and exhibited differential expression in liver tissues and 17 cell types in the liver. In addition, we reported a wide range of aging-related diseases associated with the 71 proteins, of which 12 proteins have been targeted for drug repositioning. We found that 22 genetic variants might regulate the specific protein abundance in aging, and most plasma metabolites might mediate the pathways linking the potential proteins to aging. Our findings bridge critical knowledge gaps of potential proteomic biomarker and molecular mechanisms of aging process, thereby facilitating the development of more personalized and effective therapeutic strategies against aging.

Maintaining homeostasis is crucial for successful aging, while significant disruptions in stable physiological conditions, as reflected by alterations in the proteome, may signal accelerated aging and increased disease risk [55]. Despite the various investigations of the proteomic characterization for aging-related diseases [56–60], the evidence regarding comprehensive proteome landscape of aging is largely limited. Previous proteomics studies on aging mainly centered on chronological age as the key outcome, while a single aging indicator may not fully and accurately capture the whole landscape of an individual’s aging process. By leveraging the SOMAscan proteomic platform in 1025 participants, the LonGenity cohort reported 754 of 4265 proteins associated with chronological age [16]. Using plasma proteome data from the SOMAscan assay involving 4690 participants, a previous study discovered 211 significant proteins and replicated 105 proteins linked to healthy longevity, defined as survival to 90 years of age [17]. This study expanded the list of blood-based biomarkers for human longevity that could be prioritized for translation in geroscience trials. Notably, plasma proteins can potentially be utilized in predicting chronological age and constructed proteomic clocks using a composite score of proteins selected by machine learning techniques, enabling the potential of predicting age-related diseases, multimorbidity, and mortality [12, 14, 61]. Proteomic aging clocks show great potential as biomarkers of biological age, offering the advantage of easily assessing the impact of anti-aging lifestyles and therapeutic interventions. More important, our study supplements the knowledge of the proteomic signatures of the multidimensional aging phenotypes, particularly for functional and phenotypic aging metrics, such as frailty and healthspan. Identifying the key proteomic biomarkers of functional aging will be essential for diagnosis, prognosis, and the development of new treatments for frailty symptoms and aging-related diseases aimed at improving health expectancy.

Recent advances in proteomics have significantly expanded our understanding of aging-related phenotypes. Several studies have explored specific dimensions of aging, including frailty [18], telomere length [19], healthspan [20], and age-related plasma proteome changes [21, 22]. While these investigations have identified important protein associations with individual aging phenotypes, our study distinctively addresses the multidimensional nature of aging through a comprehensive integrated analysis. For instance, Xu et al. [18] identified GDF15 among the most important proteins predicting mortality among prefrail and frail participants, while Kuo et al. incorporated various proteins in their healthspan proteomic score [20]. Similarly, Zhao et al. prioritized APOE among proteins causally associated with telomere length [19]. However, our study uniquely bridges these isolated findings through a holistic examination of multiple aging dimensions simultaneously, providing a coherent proteomic landscape that spans functional, phenotypic, and molecular aging metrics.

Our observational analyses revealed thousands of proteins associated with multidimensional aging phenotypes, which substantially exceeds findings from previous smaller-scale studies. This extensive proteomic perturbation likely reflects the systemic and interconnected nature of biological aging, where dysfunction in one system cascades to affect multiple pathways simultaneously. However, we acknowledge that many of these phenotypic associations may represent downstream consequences of aging rather than causal drivers. The aging process involves complex feedback loops where initial cellular damage triggers compensatory responses, inflammatory cascades, and metabolic adaptations, resulting in widespread proteomic changes [62].

Elucidating central pathogenic factors and intervening upon them can provide insights for biological mechanisms that delay aging and extend life expectancy. These factors could serve as promising biomarkers in clinical trials for therapies based on geroscience. The potential proteins identified in relationship with multidimensional aging phenotypes point toward potential novel aging mechanisms and pathways. Our study identified a set of proteins that were associated with aging, including UMOD, HEXIM1, ACP1, EFEMP1, ATXN2L, FURIN, CELSR2, APOE, NT5C3A, IL1RN, KIT, SERPINF2, CEP170, SERPINA1, ARFIP1, NFATC1, TRDMT1, PARP1, TK1, TYMP, RPA2, and COMMD1, known for their roles in immune response and inflammation processes. APOE is particularly noteworthy for its multifaceted role in aging biology. As a major cholesterol carrier, APOE has well-established associations with cardiovascular health and neurodegenerative pathologies, particularly Alzheimer’s disease [54]. The APOE ε4 allele is recognized as the strongest genetic risk factor for late-onset Alzheimer’s disease, accelerating amyloid-β deposition and impairing its clearance [63]. Beyond neurodegeneration, APOE influences systemic inflammation, oxidative stress, and mitochondrial function, exhibiting a critical rule in the aging process [64, 65]. In our study, we revealed various evidence of APOE in the aging biology, including the center role of the PPI network and identification of genetic variants (rs429358) regulating the protein level. These observations suggest that APOE may represent a key node connecting various molecular pathways of aging, making it a promising therapeutic target for interventions aimed at healthy aging. Notably, we found GDF15 exhibited strongly significant phenotypic associations with aging, but null genetic effects on aging. There is also considerable evidence surrounding the controversy between observational studies [12, 66, 67] and MR-identified associations [68–70] with aging-related diseases. This underscores a possible disconnect between genetic determinants of aging and proteomic signatures of aging. Alternatively, it may be that the tissue-specific expression of these proteins does not effectively translate into a blood-based aging signal.

This study has some strengths. First, based on high-throughput proteomic technologies, we employed a two-stage design to identify the potential proteomic signatures for both phenotypic and MR-identified associations, which ensured that the proteins with stable associations with aging. Second, our study integrated different plasma proteome profiling across diverse populations, including UK Biobank and FinnGen cohort. Nevertheless, our study also has several limitations. First, it is important to acknowledge that our cross-sectional proteomic analysis presents inherent limitations in capturing the dynamic nature of protein expression across the lifespan. Previous research by Lehallier et al. [21] demonstrated that plasma proteome profiles exhibit undulating patterns with age, with distinct waves of changes occurring in different decades of life. Our Mendelian randomization approach identifies proteins with genetic evidence for causality, but this does not preclude additional non-genetic regulatory mechanisms that may influence protein abundance with advancing age. Second, our analytical approach necessarily involved stringent filtering criteria to prioritize proteins with robust genetic effects on aging, which may have excluded biologically relevant proteins with weaker genetic signals or those regulated predominantly through non-genetic mechanisms. While this rigorous methodology enhances statistical confidence in our findings, it potentially overlooks proteins involved in aging processes that operate through complex epigenetic, post-translational, or environment-responsive pathways. Third, our observational analyses utilized only the Olink Explore 3072 assay from the UK Biobank, which does not capture the entire human proteome available on other platforms and panels, such as the larger Olink HT (approximately 5000 proteins) or SOMAscan (over 10,000 proteins) panels. The detection range, sensitivity, and specificity of available proteomic technologies continue to evolve, and future technological advancements may reveal additional aging-associated proteins not identified in our analysis. Furthermore, plasma proteomics primarily reflects secreted or leaked proteins from various tissues, potentially missing critical intracellular proteins involved in aging mechanisms that do not significantly alter plasma levels. These limitations underscore the need for complementary approaches integrating tissue-specific proteomics, metabolomics, and other omics data to achieve a more comprehensive understanding of the molecular landscape of aging. Finally, the absence of comparable large-scale cohorts with plasma proteome data as well as incomplete information on the multidimensional aging phenotypes limited our ability to externally validate the phenotypic associations between the identified proteins and aging. Future proteome projects are substantial to validate our findings.

Conclusions

In summary, our large-scale proteomic analysis informed a robust atlas that portrayed the proteomic landscape underpinning healthy aging. We identified proteomic signatures associated with the multidimensional aging phenotypes, and inflammation and cellular senescence may exert important biological function in the human aging. This study highlights the importance of large-scale proteomic profiling in broadening our understanding of aging biology and provides promising clues for personalized therapeutic interventions targeting aging. Our findings serve as a foundational resource on the potential biomarkers and pathophysiological precursors of aging, paving the way for the development of targeted therapeutic strategies for delaying aging and treating aging-related diseases in the future.

Abbreviations

BMI

Body mass index

CI

Confidence interval

FDR

False discovery rate

GO

Gene Ontology

GTEx

Genotype-Tissue Expression

GWAS

Genome-wide association study

ICD-10

International Classification of Diseases 10th Revision

IV

Instrumental variable

IVW

Inverse-variance weighted

KDM-BA

Klemera and Doubal’s method biological age

LD

Linkage disequilibrium

LDSC

Linkage disequilibrium score regression

LTL

Leukocyte telomere length

MHC

Major histocompatibility complex

MR

Mendelian randomization

MR-PRESSO

MR pleiotropy residual sum and outlier

MR-RAPS

MR robust adjusted profile score

MWAS

Metabolome wide association study

NMR

Nuclear magnetic resonance

PP

Posterior probabilities

PheWAS

Phenotype-wide association analyses

PPI

Protein-protein interaction

pQTL

Protein quantitative trait loci

REGENIE

Whole-genome regression tool

SD

Standard deviation

SE

Standard error

SNP

Single nucleotide polymorphism

SMR

Summary-data-based Mendelian randomization

UKB-PPP

UK Biobank Pharma Proteomics Project

Authors’ contributions

ZC and CX conceived the study. ZC and HC curated data and performed statistical analyses. ZC, HC, and JM interpreted results. ZC and HC drafted the manuscript; ZC and CX revised it for important intellectual content. CX obtained funding and provided administrative and technical support. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (No. 72204071); Zhejiang Provincial Natural Science Foundation of China (No. LY23G030005); and Scientific Research Foundation for Scholars of HZNU (No. 4265C50221204119). The funders had no role in study design, data collection and analysis, and decision to publish or preparation of the manuscript.

Data availability

Individual-level UK Biobank data are available under controlled access due to privacy and consent restrictions via the UK Biobank Access Management System (https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access). The aging GWAS summary statistics can be downloaded from Zenodo (https://zenodo.org/records/16905281). FinnGen proteomic summary results are accessible at https://www.finngen.fi/en/access_results. Parental lifespan GWAS summary statistics are available at https://datashare.ed.ac.uk/handle/10283/3209; The single-cell sequencing data from the human liver were obtained from the Gene Expression Omnibus database (GSE192742). All code used to perform the analyses are available in the GitHub Repository (https://github.com/sakuramodokich/proteomic_aging).

Declarations

Ethics approval and consent to participate

UK Biobank received ethical approval from the NHS National Research Ethics Service (North West-Haydock; REC 21/NW/0157). All participants provided informed consent at recruitment. This research has been conducted using the UK Biobank Resource under Application Number 79095. This study conforms to the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.