Atlas of genetic architecture and exploratory temporal interaction networks of plasma protein-based organ and system aging

Abstract

Organ-level investigations provide insights into inter-individual heterogeneity in health and disease, yet aging across organs and systems remains incompletely understood. Leveraging plasma proteomic data from 53,014 UK Biobank participants, we construct 16 organ- and system-specific aging clocks that show 5436 significant associations with 1059 diseases (i.e., 655 incident and 404 prevalent diseases) and widespread responses to 841 diverse environmental factors, highlighting organ-specific and pan-organ patterns. These clocks are benchmarked against two existing models and then externally validated in the China Kadoorie Biobank (n = 3977), supporting their robustness and generalizability. Genetic analyses identify 261 genomic loci associated with these clocks, revealing organ-specific and pleiotropic genetic architectures. Notably, the immune system and liver exhibit the strongest genetic connectivity with other organ systems, closely mirroring their phenotypic correlations. Using longitudinal data from 1006 participants, we apply a multitask Bayesian framework to explore potential temporal interaction networks across organs and systems, identifying temporal precedence of the immune system, liver and stomach, with sex-specific patterns characterized by stronger arterial precedence in females and stomach precedence in males. Together, this study provides a comprehensive atlas integrating genetic, environmental, and interactive insights into multiple plasma protein-based organ and system aging, informing future research on targeted anti-aging strategies.

This study provides a comprehensive atlas integrating genetic, environmental, and interactive insights into organ and system aging using multiple plasma protein-based networks, informing future research on targeted anti-aging strategies.

Introduction

Aging is a complicated and multidimensional process characterized by physiological changes that range from intracellular molecular alterations to systemic functional deterioration. Over the past decade, various organism-level aging clocks have been developed to predict an individual’s biological age, based on both clinical biomarkers and molecular omics data (e.g., Phenotypic Age and epigenetic clocks)^1–4. The discrepancy between predicted biological age and chronological age, commonly referred to as the age gap, provides a quantitative measure of aging. A positive age gap—when predicted biological age exceeds chronological age, indicating age acceleration—has been associated with elevated risks of chronic diseases, multimorbidity, and mortality^1–4. The accumulation of proteomic data over the past two years allows researchers to study the aging process across multiple organs and systems by developing organ- and system-specific clocks, representing a paradigm shift from organismal aging to organ and system aging⁵. Both environmental exposures and genetics are known to play important roles in shaping the human aging process. However, little is known about the genetic architecture of multiple organ and system aging derived from proteomics⁶. Despite emerging epidemiological evidence linking environmental exposures to organ and system aging, the sparse coverage of environmental exposures limits understanding of these complicated associations⁷. In summary, a gap remains in our understanding of the full spectrum of genetic and environmental influences on multiple organ and system aging. Bridging this knowledge gap may provide potential therapeutic targets and inform the development of prevention strategies to potentially slow organ and system aging, ultimately reducing the risk of chronic diseases and promoting healthy longevity.

Both in vitro and in vivo studies have demonstrated that aging can be transmitted in a paracrine manner within affected organs^8–11. Recent experimental studies in mice have suggested that certain organs may play a coordinating role in the overall aging of the organism. Yet, whether aging can spread systemically to distant organs, or specific organs exhibit temporal precedence in systemic aging in humans, remains incompletely understood. Moreover, previous statistical methods (e.g., Fast Greedy Equivalence Search) impose stringent topological and identifiability constraints¹². This inherent structural rigidity severely restricts their generalizability in large-scale causal discovery within complex biological networks that feature feedback loops, cycles, or dynamic mutual dependencies. Another unanswered question is whether there are potential sex- and age-differences in this regard.

Here, leveraging the largest proteome-phenome dataset to date from the UK Biobank (UKB), following published procedures^5,6,13–16, we calculated biological aging clocks for 16 organs and systems based on 2923 plasma proteins, and delineated their associations with 1059 diseases (the most comprehensive disease atlas to date), including the onset of 655 diseases and the prevalence of 404 diseases in 53,014 participants. External validation was further conducted in an independent cohort, the China Kadoorie Biobank (CKB; n = 3977). We then characterized the responses of these aging clocks to 841 environmental factors (e.g., lifestyle factors, physical measurements, and living environment). Furthermore, we identified the genetic architecture (i.e., genomic loci and genetic heritability) and biological pathways of aging across organs and systems through integrating genome data. Finally, based on longitudinal data from three assessments of 1006 participants, we used a multitask Granger-like Bayesian regression framework with Automatic Relevance Determination (ARD, see “Methods”) to explore potential temporal interaction networks underlying multiple organ and system aging. This panoramic atlas serves as a resource for describing the heterogeneous associations of multiple organ and system aging with hundreds of disease risks and exposures, and also their genetic basis.

Results

Calculations of organ- and system-specific aging clocks

Our primary study population included 53,014 UKB participants (aged 56.4 ± 8.2 years, 52.3% female) with available plasma proteomic data at baseline. To develop organ- and system-specific aging clocks, these participants were randomly split into a training (n = 37,110) and a testing set (n = 15,904) in a 7:3 ratio (Fig. 1a).

Fig. 1. Aging clocks of organs and systems constructed using plasma proteins.

a Study populations for constructing organ- and system-specific aging clocks. Plasma proteomic data from the UK Biobank (UKB) were collected at three distinct time points: 2006–2010, 2014+, and 2019+. For cross-sectional analyses, 53,014 participants with available proteomic data from the initial assessment period (2006–2010) were included and randomly divided into training (70%) and testing (30%) sets to develop and validate organ- and system-specific aging clocks. For longitudinal analyses, a subset of 1006 participants with available proteomic data across all three time points were included and similarly divided. b Enrichment of proteins to specific organs and systems. A gene was considered enriched in a particular organ or system if its expression level was at least fourfold higher compared to any other, as determined using the GTEx bulk organ RNA-seq database. A total of 631 proteins meeting this criterion in the UKB were used to construct organ- and system-specific aging models. c Performance and predominant proteins contributing to heart and brain models. Shapley values were calculated for each organ- and system- enriched protein to quantify the relative importance to their corresponding aging models. d Bar plots showing Pearson correlation coefficients (top) and mean absolute error (MAE) (bottom), quantifying the accuracy of age prediction for each organ and system. e Clusters of extreme agers. Extreme agers were identified as individuals exhibiting age gaps exceeding two standard deviations above the population mean for at least one organ or system. f The mean age gaps of organs and systems among heart extreme agers, brain extreme agers, and liver extreme agers. Source data are provided as a Source Data file. The figure was created in BioRender. Pu, F. (2026) https://BioRender.com/6w5xmne.

Following the methodology by Oh et al.⁵, we mapped the putative organ- and system-specific plasma proteome by defining a protein as organ- and system-specific if its corresponding RNA expression in the Genotype-Tissue Expression (GTEx) database was at least fourfold higher in one organ compared to any others (“Methods”; Fig. 1b and Supplementary Data 1). Among the 2923 proteins available from the UKB, 631 (18%) met this criterion, with the immune system contributing the most proteins (n = 271) and the thyroid the fewest (n = 2) (Supplementary Data 2). Using these mutually exclusive organ- and system-enriched proteins as input, we developed 18 organ and system-specific aging models via Extreme Gradient Boosting (XGBoost) in the training set (“Methods”). After excluding two models (i.e., adrenal and thyroid) due to their low correlation with chronological age in the testing set (Pearson r < 0.3), we retained 16 robust organ- and system-specific aging models for downstream analyses, including previously unreported models for the salivary gland, pituitary, and skin¹². All models demonstrated strong predictive accuracy for chronological age in the testing set. For example, the artery and brain models achieved a mean absolute error (MAE) of 4.16 and 4.12 years, respectively (Fig. 1c, d). Correlation and accuracy metrics for the other aging models are provided in Supplementary Fig. 1.

Age gap of each organ and system was then estimated as the participant’s biological age relative to same-aged peers⁴ (“Methods”; Supplementary Data 3). These age gaps showed low-to-moderate correlations across organs and systems, with pairwise Pearson r ranging from 0.006 to 0.325 (Supplementary Data 4), indicating substantial heterogeneity in aging. Further, we identified 8324 participants (15.7%) whose age gaps exceeded two standard deviations (SD) above the population mean for at least one organ or system as extreme agers, and subsequently clustered them into distinct extreme aging types (Fig. 1e). Consistent with findings by Oh et al.⁵, most extreme agers exhibited accelerated aging in only one organ or system. For example, over 21% of extreme agers were clustered into the heart extreme aging type; their age gaps of heart averaged 2.51 SD above the population mean, while age gaps of other organs and systems were marginally above zero (Fig. 1f). Only two clusters presented co-occurring extreme aging in more than one organ: cluster 11 exhibited extreme aging in the brain and immune system, and cluster 14 exhibited extreme aging in the esophagus and stomach.

To evaluate the cross-cohort generalizability of our organ-specific aging clocks, we directly applied the UKB-trained models to an independent cohort of 3977 participants (54% female, age range 30–78 years) from the CKB. Across the 16 organs included in downstream analyses, predicted organ biological ages correlated robustly with chronological age in CKB, with Pearson r values ranging from 0.14 to 0.71 (e.g., brain: r = 0.71; artery: r = 0.69; immune: r = 0.64). Notably, organ-level predictive performance in CKB was largely concordant with that in the UKB testing set (cross-cohort Pearson r = 0.844, P < 0.001), with minimal attenuation upon cross-cohort transfer (e.g., brain: r = 0.77 in UKB vs. r = 0.71 in CKB; artery: 0.77 in UKB vs. 0.69 in CKB; immune: 0.70 in UKB vs. 0.64 in CKB) (Supplementary Data 5). Furthermore, to benchmark our organ-specific aging clocks against existing frameworks under identical conditions, we applied two established proteomic aging models^13,16 to the same UKB held-out testing set. We found that our model showed slightly better performance than both the Goeminne¹³ and Oh¹⁶ models across most organs examined (e.g., brain: r = 0.77 vs. 0.73 [Goeminne] and 0.71 [Oh]; immune: 0.70 vs. 0.55 [Goeminne] and 0.55 [Oh]; kidney: 0.47 vs. 0.42 [Goeminne] and 0.18 [Oh]), as shown in Supplementary Fig. 2, with full results provided in Supplementary Data 5.

Associations of multiple organ and system aging with 1059 diseases

We examined the associations between age gaps of multiple organs and systems and the risk of 655 incident diseases and 404 prevalent diseases using Cox proportional hazards regression and logistic regression models, respectively (“Methods”; Supplementary Data 6, 7). After applying a stringent Bonferroni correction, we identified 3387 significant associations between age gaps and incident diseases (P < 0.05/[16 × 655] = 4.77 × 10⁻⁶) and 2049 significant associations with prevalent diseases (P < 0.05/[16 × 404] = 7.74 × 10⁻⁶) (Fig. 2a, b, Supplementary Data 8 and 9). The majority of these associations were consistent across both incident and prevalent disease analyses (Fig. 2c). Notably, age gaps exhibited distinct disease association patterns across organs and systems. Specifically, organs and systems primarily affected by a given disease pathology generally show the strongest associations and largest effect sizes for diseases within that category. For instance, the brain age gap demonstrated the highest hazard ratio (HR) for vascular dementia, followed by Alzheimer’s disease. The lung age gap strongly predicted idiopathic pulmonary fibrosis (IPF) and interstitial lung disease (ILD), whereas the kidney age gap best predicted hypertensive renal disease and dialysis. Similarly, for prevalent diseases, the heart age gap showed the highest odds ratios (OR) for atrial fibrillation and flutter, and cardiomyopathy, whereas the lung age gap again demonstrated strong associations with IPF and ILD. All 16 organ age gaps were significantly associated with all-cause mortality risk (5765 deaths), with kidney and brain aging showing the strongest associations (HR = 1.53, 95% CI: 1.48–1.58, and HR = 1.48, 95% CI: 1.43–1.52, respectively) (Supplementary Data 10). Subgroup analysis stratified by sex revealed that more than half of the associations remained robust (Fig. 2d and Supplementary Data 11–14). Notably, certain diseases, such as kidney-related disorders (e.g., dialysis, chronic kidney disease, renal tubulointerstitial diseases, and acute renal failure), obesity, and hypertension, demonstrated extensive associations with age gaps across multiple organs and systems, which were consistent with previous reports^12,17. This pattern may reflect either the profound impact of renal pathophysiology on circulatory proteomic profiles^18,19, or the secondary multi-organ functional decline resulting from renal dysfunction. In the sensitivity analysis, when excluding participants developing diseases within the first 2-year period, we identified 3165 significant associations between age gaps and incident diseases (618 incident diseases) (P < 0.05/[16 × 618] = 5.06 × 10⁻⁶, Supplementary Data 15). The vast majority of these associations (99.4%) were directionally consistent with the main analysis, and notably, no significant associations in the opposite direction were observed.

Fig. 2. Disease atlas of multiple organ and system aging.

a, b. Associations between organ- and system-specific age gaps and 655 incident (a) or 404 prevalent (b) diseases, identified using Cox (a) and logistic (b) regression models. Only statistically significant associations are plotted (3387 associations for incident diseases and 2049 for prevalent diseases). Each point represents a significant age gap-disease association, colored by ICD-10 disease chapter. c Comparisons of significant associations in prevalent versus incident diseases. Dark blue bars represent significant associations unique to prevalent diseases; yellow bars represent those unique to incident diseases; purple bars represent associations common to both. d Comparison of significant associations in the sex subgroup analysis (left: incident diseases; right: prevalent diseases). e Selected significant associations between extreme ager clusters and incident diseases. Bar colors correspond to different extreme ager clusters. Proportions of each cluster within the overall population are shown in brackets. Source data are provided as a Source Data file.

We further examined associations between 16 organ age gaps and 109 incident diseases using Cox proportional hazards models in the CKB. Consistent with findings in UK Biobank, the kidney age gap showed broad and strong associations with incident diseases, including type 2 diabetes (HR = 2.13, 95% CI: 1.92–2.37), essential hypertension (HR = 1.34, 95% CI: 1.27–1.43), and renal failure (HR = 2.16, 95% CI: 1.61–2.91). Heart age gap was associated with cardiac arrhythmia (HR = 1.21, 95% CI: 1.06–1.39) and heart failure (HR = 1.39, 95% CI: 1.20–1.60) (Supplementary Data 16). Accelerated aging of multiple organs was significantly associated with elevated all-cause mortality risk (1426 deaths), with the strongest associations observed for pancreas (HR = 1.60, 95% CI: 1.38–1.86), kidney (HR = 1.24, 95% CI: 1.05–1.45), and intestine (HR = 1.28, 95% CI: 1.11–1.47) (Supplementary Data 17). Together with previous observations, these results suggest good generalizability of our organ-specific aging clocks across diverse populations and demonstrate that these organ age gaps exhibited prognostic relevance in an independent East Asian population.

Moreover, we investigated the associations between participants with the extreme aging types identified above and 437 incident diseases (excluded diseases with incident cases <100), revealing both remarkable organ specificity and unprecedented effect magnitudes (Fig. 2e and Supplementary Data 18). After Bonferroni correction, we identified 725 significant associations. Participants with the heart extreme aging type showed a broad disease spectrum (169 incident diseases). Notably, participants with extreme aging types exhibited a substantially increased risk of incident diseases specifically related to their corresponding organs or systems. For example, participants with the kidney extreme aging type had a 17.0-fold increased risk of chronic kidney disease, those with the heart extreme aging type had a 15.3-fold increased risk of cardiomyopathy, and those with the brain extreme aging type had a 7.1-fold increased risk of unspecific neurodegenerative disorder. Particularly, participants with the multi-organ extreme aging type demonstrated severe systemic adverse health consequences, including a 10.8-fold increased risk of acute renal failure, a 7.1-fold increased risk of pneumonia, and a 5.83-fold increased risk of type 2 diabetes. The extraordinarily high risk of incident diseases observed in participants with extreme aging phenotypes underscores the urgent need to identify and intervene in these high-risk subpopulations.

Multiple organ and system aging broadly responds to environmental factors

We next examined the associations between 841 environmental factors and age gaps of multiple organs and systems using linear regression models, identifying 6430 significant associations after Bonferroni correction (P < 0.05/[16 × 841] = 3.72 × 10⁻⁶) (Fig. 3, Supplementary Data 19 and 20). Several key factors, such as physical activity, socioeconomic status and physical measurements, demonstrated widespread protective or detrimental effects on multiple organs and systems aging. For example, participants who reported extremely liking (vs. extremely disliking) taking the stairs showed significantly reduced age gaps in ten organs and systems (e.g., adipose: β = −0.37, immune: β = −0.35, and kidney: β = −0.40). Similarly, participants with higher socioeconomic status, such as those who were employed (vs. unemployed) or had higher average total household income before tax (e.g., greater than £100,000 vs. less than £18,000), showed slower aging across nearly all organs and systems. Conversely, smoking status showed strong associations with accelerated organ aging, with current smokers showing significantly larger age gaps in 13 organs and systems (e.g., stomach: β = 0.63, brain: β = 0.34, and skin: β = 0.40). Notably, former smokers and participants exposed to maternal smoking during early life or secondhand smoke (tobacco smoke at/outside the home) also showed significantly accelerated aging in nine and eight organs and systems, respectively.

Fig. 3. Environmental atlas of multiple organ and system aging.

Associations of environmental factors (e.g., lifestyle factors, physical measurements, and living environment) with age gaps across organs and systems, colored by each organ and system. Only significant associations are plotted (n = 6430), and associations with coefficient sizes in the top 50% of each direction (positive/negative) are colored. Abbreviations: Smoking (Most/All Days): Tobacco smoking (Smokes on most or all days); Illness/Injury/Stress (Serious illness/injury to self): Illness, injury, bereavement, stress in last 2 years (Serious illness, injury or assault to yourself); MUFA/TFA%: Monounsaturated Fatty Acids to Total Fatty Acids percentage; PUFA/MUFA: Polyunsaturated Fatty Acids to Monounsaturated Fatty Acids ratio; Income (>100 K): Average total household income before tax (Greater than 100,000); Vigorous PA (6d/week): Number of days/week of vigorous physical activity 10+ min (6 days); Alcohol Frequency (Daily/Almost daily): Alcohol intake frequency (daily or almost daily); Stair Liking (Extremely): Liking for taking the stairs (extremely like); Stair Liking (Neutral): Liking for taking the stairs (neither like nor dislike); PA Type, 4wk (Strenuous sports): Types of physical activity in last 4 weeks (Strenuous sports); Shortness of Breath (Bothered a lot): Degree bothered by shortness of breath in the last 3 months (Bothered a lot). Source data are provided as a Source Data file. The figure was created in BioRender. Pu, F. (2026) https://BioRender.com/djxwy6q.

Interestingly, metabolic factors exhibited broad detrimental or protective effects across multiple organs and systems. Higher fat-related factors (e.g., leg fat percentage, arm fat percentage, and trunk fat percentage) were widely associated with larger age gaps across organs and systems. For example, a 1-SD increase in leg fat percentage was significantly associated with accelerated aging in 12 organs and systems (e.g., adipose: β = 0.56, esophagus: β = 0.41, and kidney: β = 0.44). Metabolomic profiling further revealed a strong association pattern between lipid biomarkers and aging across organs and systems: most high-density lipoprotein (HDL)-related biomarkers (e.g., total lipids, large HDL, and HDL cholesterol) were associated with smaller age gaps, whereas very low- and low-density lipoprotein-related biomarkers were associated with larger age gaps.

Meanwhile, aging across multiple organs and systems exhibits heterogeneous responses to certain factors—such as dietary habits. For instance, preference for abyssal fishes like salmon (β = −0.17) and mackerel (β = −0.11) was significantly associated with a reduced intestine age gap. These effects may be attributed to the anti-inflammatory properties and relatively high omega-3 fatty acid content of these fish, which are known to promote cellular health. In contrast, excessive consumption of ultra-processed meat products, such as bacon (β = 0.27) and burgers (β = 0.20), was significantly associated with a larger kidney age gap, possibly mediated by elevated oxidative stress and inflammation, which can drive cellular senescence and tissue dysfunction. Further subgroup analysis stratified by sex revealed that more than half of the associations remained robust (Supplementary Data 21 and 22).

Genetic architecture of multiple organ and system aging

To investigate the genetic mechanisms underlying aging across multiple organs and systems, we conducted genome-wide association studies (GWAS) of age gaps across 16 organs and systems (“Methods”). We identified 26,985 significant SNPs associated with age gaps (P < 5 × 10⁻⁸) and a total of 261 genomic loci (defined as a ± 250-kb region around the genome-wide significant SNPs) significantly associated with age gaps of these organs and systems were identified, harboring 760 lead SNPs (Supplementary Data 23). SNP-based heritability (h²) estimates were statistically significant for all organs and systems except the heart and pituitary. The esophagus exhibited the highest heritability (h² = 0.46), followed by the kidney (h² = 0.42), adipose (h² = 0.37), and immune system (h² = 0.36) (Supplementary Data 24). We further estimated the genetic correlation (g_c; “Methods”) among age gaps of organs and systems (Fig. 4a and Supplementary Data 25). After Bonferroni correction, the immune system, liver, and lung age gaps showed the greatest number of significant genetic correlations with other organs and systems. Notably, age gaps of liver and artery showed the highest genetic correlation (g_c = 0.52), which corresponds to the relatively high phenotypic correlation (Pearson r = 0.30 in Fig. 1c).

Fig. 4. Genetic atlas of multiple organ and system aging.

Genetic analyses were based on GWAS summary statistics for organ system age gaps from n = 53,014 UKB participants of European ancestry after genetic quality control. For panels c–e, each bar represents one organ system age gap–annotation pair, and error bars indicate standard errors estimated by stratified LD score regression. a A genetic correlation between 16 organ and system age gaps. The color scale represents the strength of the correlation, ranging from negative (red) to positive (blue). ***P < 0.001, **P < 0.01, *P < 0.05. Exact P values are provided in Supplementary Data 25 and the Source Data file. b Gene-level biological pathway annotation. Gene set enrichment analysis (GSEA) was performed using curated gene sets and Gene Ontology terms from the MSigDB database. c Tissue-specific partitioned heritability estimates using gene sets from multi-tissue gene expression data. d Partitioned heritability estimates for the general 53 functional categories. For visualization purposes, we only show the four categories with the highest significant estimates for each age gap of organs and systems. The label for 500 denotes a 500-base pair window around each of the 24 main annotations in the full baseline model, which prevents a biased estimate inflated by heritability in flanking regions. e Tissue- and chromatin-specific partitioned heritability estimates using multi-tissue chromatin data. f Gene-drug-diseases network constructed using GREP, only significant results (P < 0.05, one-sided Fisher exact test) are shown. Detailed data are provided in Supplementary Data  32. Source data are provided as a Source Data file. The figure was created in BioRender. Pu, F. (2026) https://BioRender.com/huyq3t3.

Next, we performed gene-based association analyses using the MAGMA software, followed by gene set enrichment analysis (GSEA) of significantly associated genes to elucidate the biological pathways underlying the aging process of each organ and system. Genes associated with age gaps in each organ and system were enriched in multiple biologically relevant pathways corresponding to the specific functions and aging mechanisms of that system (Fig. 4b, Supplementary Data 26 and 27). For example, genes associated with the adipose age gap were significantly enriched in the lipid transporter activity pathway (P = 5 × 10⁻⁷). Genes associated with the artery age gap were implicated in molecules associated with elastic fibers, which corresponds to elastin being the predominant contributor to the artery aging clock. Meanwhile, genes associated with the immune system age gap showed enrichment in multiple immune-related pathways, including immunoregulatory interactions between a lymphoid and a non-lymphoid cell (P = 1.9 × 10⁻¹⁰), immune response (P = 1.1 × 10⁻⁸), and positive regulation of immune system process (P = 5.5 × 10⁻⁸).

Furthermore, we performed partitioned heritability analyses to estimate the heritability enrichment of genetic variants associated with age gaps of organs and systems, focusing on annotations related to three different cell types (i.e., neurons, oligodendrocytes, and astrocytes), 205 tissue-specific gene expression data, 53 non-tissue-specific functional categories, and 489 tissue-specific chromatin states (“Methods”). We found significant heritability enrichment in neurons for the pancreas and adipose age gaps (both P = 0.02, Supplementary Data 28). Heritability enrichment in the tissue-specific gene expression revealed that the organ and system age gaps were not only enriched in their corresponding tissues but also in others, further indicating cross-organ sharing of genetic mechanisms underlying aging (Fig. 4c and Supplementary Data 29). For example, the heart age gap showed significant heritability enrichment in multiple tissues, including the heart atria (P = 1.3 × 10⁻³), atrial appendage (P = 2.0 × 10⁻³), and kidney (P = 8.1 × 10⁻³). Similarly, the immune age gap showed expected enrichment in the blood (P = 2.4 × 10⁻⁴), neutrophils (P = 3.1 × 10⁻⁴), and bone marrow cells (P = 1.5 × 10⁻³), as well as in non-immune tissues such as the stomach (P = 2.1 × 10⁻³) and rectum (P = 4.2 × 10⁻³). Concerning heritability enrichment in non-tissue-specific functional categories, repressed genomic regions, histone H3 acetylated lysine 27 (H3K27ac) regions, and super-enhancer regions showed the strongest heritability enrichment across multiple organ systems (Fig. 4d and Supplementary Data 30). For example, the immune (P = 1.9 × 10⁻⁷), kidney (P = 4.5 × 10⁻⁷), liver (P = 1.2 × 10⁻⁴), lung (P = 5.3 × 10⁻⁹), and salivary (P = 4.8 × 10⁻⁵) age gaps exhibited the highest heritability enrichment in repressed regions. Super-enhancers are typically characterized by high-density H3K27ac modifications. The H3K27ac regions, which are associated with higher transcriptional activation, used 38.9% of SNPs to explain the SNP heritability of age gaps in multiple organs and systems. Notable examples included the adipose (enrichment = 2.26 ± 0.30, P = 2.46 × 10⁻⁵), artery (enrichment = 2.41 ± 0.28, P = 2.1 × 10⁻⁷), and heart (enrichment = 2.12 ± 0.32, P = 6.1 × 10⁻⁴) age gaps. In addition to H3K27ac, enrichment analyses of tissue-specific chromatin states revealed that the heritability of age gaps across organs and systems was enriched in more transcriptional activation-associated histone modifications (e.g., H3K4me1 and H3K4me3) within their respective tissues (Fig. 4e and Supplementary Data 31). For example, the heritability of the brain age gap was significantly enriched in H3K4me3-marked regions of the middle hippocampus (P = 1.5 × 10⁻⁴) and angular gyrus (P = 3.2 × 10⁻³), both functionally relevant brain regions. Moreover, certain organ systems age gaps showed enrichment in H3K36me3-marked regions, which are tightly located in actively transcribed genome regions, such as the brain cingulate gyrus (P = 1.1 × 10⁻³) for the brain age gap and primary monocytes from peripheral blood (P = 5.0 × 10⁻⁴) for the immune age gap. These findings suggested that the regulation of gene expression, through both activation and repression, constitutes a key genetic mechanism underlying organismal aging, while distinct regulatory elements such as promoters, enhancers, and gene bodies may contribute uniquely to the aging process of different organs and systems.

To assess whether organ aging-associated genetic variants implicate druggable targets, we mapped lead SNPs to genes using Functional Mapping and Annotation (FUMA) with two complementary strategies: positional mapping (10-kb window) and expression quantitative trait loci (eQTL) mapping based on GTEx v8 data. This yielded 1069 positional genes and 1496 eQTL-mapped genes, totaling 1762 unique genes for drug target enrichment analysis using Gene Repurposing using Expression Profiles (GREP)²⁰. We identified 16 target genes linked to drugs across four organ systems (Fig. 4f and Supplementary Data 32). For adipose aging, peroxisome proliferator-activated receptor gamma (PPARG) emerged as a key target of telmisartan and bezafibrate (cardiovascular drugs, P = 0.014). Notably, telmisartan is unique among angiotensin receptor blockers in functioning as a partial PPARγ agonist, and preclinical studies have demonstrated its ability to delay adipose tissue aging and improve insulin sensitivity through PPARγ-dependent pathways^21,22. For arterial aging, PLG (plasminogen) was linked to thrombolytic agents including alteplase and tenecteplase (P = 0.017), consistent with the established age-related decline in fibrinolytic capacity²³. These findings suggest organ-specific drug repurposing opportunities for these targets with the additional aim of mitigating aging-related decline.

Complex potential temporal interaction networks across multiple organ and system aging based on longitudinal analyses

Proteomic data were available for 1006 participants at their third and fourth visits. For longitudinal analyses, we focused on 1463 proteins measured across all three visits (baseline, third, and fourth, mean follow-up: 12.4 years), and re-trained the 16 aging models (hereafter referred to as longitudinal models) (“Methods”; Fig. 1a). The longitudinal models demonstrated comparable accuracy to the original models (Supplementary Data 33). All participants exhibited progressive biological aging across all organs and systems over time. Participants with larger age gaps at baseline tended to experience faster progression (Fig. 5a, Supplementary Fig. 3 and Supplementary Data 34). Notably, males showed significantly faster aging in eight organs and systems, particularly in the brain and liver, whereas females showed faster aging only in the pituitary (Fig. 5b and Supplementary Data 35).

Fig. 5. Complex potential temporal interaction networks of organ and system aging.

a Comparison of the aging rate of the stomach among participants with different baseline age gaps. Thick black lines represent median values, the boxes span 50% of the data, and the black whiskers span 1.5 times the interquartile range or the range of the data, whichever is smaller. Pairwise comparisons were performed using two-sided two-sample t-tests. ***P < 0.001, **P < 0.01, *P < 0.05. Exact P values are provided in Supplementary Data 34 and the Source Data file. b Comparison of the aging rates among organ systems between sexes. c Temporal interaction networks across various organs and systems among different sex. The bar plots illustrate the cumulative temporal associations among various organ systems. The red bar denotes the total number of temporal associations where an organ system shows temporal precedence over others, and the blue bars signify the total number of temporal associations where an organ system is preceded by others. The intensity of the color corresponds to the sum of the coefficients for all corresponding temporal associations. d, e. The temporal interaction networks between organ systems among the overall population (d) and females (e). The thickness of the arrow represents the strength of the temporal predictive relationship. Arrow colors denote the organ of origin: in panel d, green, blue, and red arrows indicate temporal relationships originating from the stomach, immune system, and liver, respectively; in panel e, green, red, and blue arrows indicate those originating from the stomach, artery, and immune system, respectively. Black arrows represent temporal relationships originating from other organ systems. Source data are provided as a Source Data file. d was created in BioRender. Pu, F. (2026) https://BioRender.com/1n74ggh. e was created in BioRender. Pu, F. (2026) https://BioRender.com/3lxj7z5.

To further elucidate the potential temporal interaction networks of aging across organs and systems, we employed a multitask Granger-like Bayesian regression framework with Automatic Relevance Determination (ARD) (“Methods”). In the overall longitudinal dataset, the immune system, liver, and stomach showed the strongest temporal precedence, with their aging predicting subsequent aging in other organs (Fig. 5c and Supplementary Data 36). Notably, several organ-system pairs exhibited bidirectional temporal relationships, including brain-immune system, brain-liver, liver-immune system, and immune system-lung (Fig. 5d). Specifically, a 1-SD increase in immune system age gap was temporally associated with a 0.13-SD increase in subsequent lung aging, while a 1-SD increase in lung age gap was reciprocally associated with a 0.11-SD increase in subsequent immune aging. These findings underscore the intricate interaction across organs and systems during the aging process.

Furthermore, sex-specific differences were evident in the potential temporal interaction network. While the immune system and liver showed strong temporal precedence in both sexes, the artery exhibited stronger temporal precedence over other organ systems among females (Fig. 5c, e and Supplementary Data 37), and the stomach exhibited stronger temporal precedence among males (Supplementary Fig. 4 and Supplementary Data 38).

We next investigated the potential temporal aging networks across organs and systems in different life stages. As shown in Supplementary Fig. 5 and Supplementary Data 39, the immune system consistently showed temporal precedence over other organ systems throughout the lifespan. The liver exhibited an age-dependent pattern of temporal predictive relationships, with the strongest predictive effect observed during the 40–49 and 70–79 age ranges, but a notably attenuated effect between 50 and 69 years. This suggests that the liver may play functionally divergent roles across the lifespan, potentially reflecting shifts in metabolic regulation, hormonal signaling, or inflammation at different life stages. Notably, during the 70–79 age range, temporal relationships were markedly intensified, with the liver, immune system, lung, artery, and brain showing particularly strong predictive relationships with other organ systems. The heightened interconnectivity may represent a critical transition toward multi-morbidity and functional decline, highlighting the importance of timely, age-specific interventions targeting key organs and systems showing early aging changes to delay systemic deterioration.

Discussion

This study, to the best of our knowledge, depicts the most comprehensive landscape of multiple organ and system aging in disease. By constructing protein-based aging clocks for 16 different organs and systems in 53,014 participants, this study revealed 5436 significant associations of organ and system aging with diseases, as well as widespread responsiveness of these aging clocks to diverse environmental factors (e.g., obesity-related factors, lipid metabolism factors, and smoking). Furthermore, this study unraveled the genetic architecture of the aging process across 16 organs and systems, revealing their shared genetic mechanisms and heterogeneities. Finally, leveraging longitudinal proteomic data, we explored potential temporal network patterns across multiple organ and system aging, identifying temporal patterns wherein the immune system, liver, and stomach aging preceded aging in other organ systems throughout the lifespan, although more validations with a larger sample size are needed.

Previous research has primarily explored the health consequences of organ and system aging, yet these studies have been relatively limited in scope, predominantly focusing on specific major chronic diseases such as stroke, cancer, and dementia^5,12,24. Moreover, despite the substantial contribution of the environmental exposome to aging^7,25, there is still a lack of a refined atlas detailing its associations with organ and system aging. This study stands out by linking organ and system aging to a comprehensive set of diseases and health-related phenotypes. The organ- and system-specific disease associations were systematically elucidated, revealing that the increased risk of prevalent or incident diseases was more significantly associated with the accelerated aging of the organ or system corresponding to the disease pathology. The organ systems-specific associations may be partially explained by the differential protein enrichment across organ systems. For example, LAMP3 and CCL18—two key contributors of the lung aging clock—exhibit mechanistic relevance to pulmonary function. LAMP3 critically maintains alveolar integrity by reducing gas-liquid interfacial tension in alveolar type II cells²⁶, while CCL18, a lung-predominant inflammatory protein, shows elevated serum levels in IPF²⁷ and chronic obstructive pulmonary disease²⁸. These molecular underpinnings provided support for the better predictive and diagnostic efficiency of the lung age gap for pulmonary diseases. Consistent with previous research, elevated obesity-related factors²⁹—both anthropometric and lipid metabolism indicators—and smoking³⁰ are significantly associated with accelerated aging across nearly all organs and systems. These findings provide a minimally invasive approach capable of assessing and monitoring disease risks over time, which only requires a small blood sample. The health atlas underscores the central contribution of metabolic dysregulation to organismal aging, highlighting the importance of metabolic management in delaying aging and reducing the risk of age-related multimorbidity. It also highlights the potential of managing modifiable, pan-organ-associated risk factors (e.g., smoking, physical activity) on organismal aging for primary prevention.

To our knowledge, this study represents one of the largest-scale GWAS on multiple organs and systems aging quantified by plasma proteomics, revealing the heterogeneous genetic architectures. One comparable study has explored the genetic mechanisms underlying nine organs and systems aging clocks constructed from anthropometric functional measures³¹. In contrast, aging clocks building on plasma proteomics are biologically driven, have relatively high organ specificity, and allow for a more timely and precise detection of subtle aging-related changes⁵. These protein-based aging clocks are capable of capturing more organs and systems, such as the pituitary and esophagus, which are challenging to measure accurately using anthropometric functional measures. Our analyses revealed that the heritability of aging of each organ and system is predominantly enriched in the respective tissues and biological pathways. Notably, the liver and artery exhibit a relatively high heritability correlation and share overlapping genetically enriched regions, which corresponds with their high correlation in predictive age and prior evidence linking liver conditions and cardiovascular diseases^32,33. While aging-associated cellular population dynamics have been elucidated in mice³⁴, this study provides a panoramic view of the genetic architecture of organism aging on organ levels. It lays the groundwork for future research using single-cell genomics data to explore how aging and functional deterioration in different cell types contribute to organ-level aging.

This study explored the potential temporal interaction networks among multiple organ systems, going beyond analyses of specific interorgan axes (e.g., heart-brain-liver³⁵, kidney-lung³⁶, and brain-gut³⁷). Our ARD framework leverages temporal information, prunes non-informative predictors without explicit structure search, allows reciprocal influences, is computationally efficient, and scalable, offering a flexible foundation for exploring temporal dependencies in complex biological networks. We found that the immune system showed consistent temporal precedence over other organ systems and engaged in extensive bidirectional temporal relationships in all analyses. Interestingly, Yousefzadeh et al. recently constructed a rat model of isolated immune system senescence and demonstrated that aged immune cells can drive systemic aging through a combination of cell-autonomous (loss-of-function) and cell non-autonomous (gain-of-function, such as SASP) mechanisms³⁸. Rapamycin, which extends lifespan in multiple model organisms including nematodes, fruit flies, and mice^39–41, has been shown to block age-related hyperactivation of the NF-κB-like immune deficiency pathway, thereby reducing inflammaging^38,41,42. The aged and senescent immune system appears to play a critical role in organismal aging, suggesting that immune-related targets could be a promising direction for anti-aging therapies. The liver also showed dominant temporal precedence in both sexes. Supporting this, Kiourtis et al. constructed genetic models of hepatocyte-specific senescence and found that senescence of liver cells can propagate systemically through the diffusion of transforming growth factors, leading to failure and dysfunction of other organs⁴³. Additionally, the artery demonstrates stronger temporal precedence in females, while in males, the stomach and intestine exhibited stronger temporal precedence over aging in other systems. Although the precise biological mechanisms through which these organ systems propel systemic aging remain to be elucidated, the sex difference within these organ systems during aging has been partially unveiled by previous studies. The substantial hormonal fluctuations during menopause, notably the diminished vascular protection from endogenous estrogen in premenopausal females⁴⁴, could partially account for the unique role of arteries in female aging. Regan et al. indicate that male enterocytes have higher basal autophagy levels and larger intestinal cells⁴⁵. In short, this study may shed light on the sex-specific nature of intrinsic aging mechanisms and emphasizes the need to fully consider sexual dimorphism when developing future anti-aging therapeutic strategies.

Several limitations of this study warrant attention. First, identifying organ- and system-specific plasma proteins may not comprehensively reflect the aging of the organ system (e.g., mRNA-protein correlations typically fall below 40%⁴⁶), though this remains widely used and possibly the only feasible approach to track organ health. Furthermore, the plasma proteome predominantly captures secreted proteins, with limited detection of intracellular and extracellular membrane proteins. Second, participants with certain disease conditions may introduce a slight bias when training the organ and system aging clocks. Nevertheless, the substantial sample size of the UKB provides exceptional statistical power, and the models demonstrate strong predictive performance. Third, although the UKB offers one of the largest longitudinal datasets, the repeated proteomic measurements remain limited in number, and the age range of the population is relatively narrow, approximately between 45 and 80 years, with fewer individuals at both extremes. Future studies should strive to build cohorts with broader age ranges and more comprehensive biomarkers to capture a more complete trajectory of organ and system aging throughout the entire life course, especially in early life stages, which involve rapid organismal development and high sensitivity to external exposures. Fourth, diagnoses of certain prevalent diseases may precede the collection of baseline plasma proteomic data, thus necessitating cautious interpretation of results. Fifth, considering the potential for occult disease progression and the relative delay in diagnosis, biases arising from potential reverse causality and circularity should be considered. Sixth, although we employed the Bonferroni correction to address multiple testing, the large number of associations examined may still pose a risk of false positives and inflated hazard ratios, requiring further validation. Lastly, this study was largely based on a UK Caucasian middle-aged and young-old population. Although we externally validated the organ-specific aging clocks in an independent East Asian cohort, demonstrating cross-population generalizability, future studies should further assess the generalizability of findings in more ethnically and geographically diverse populations to bolster reliability⁴⁷.

In summary, after constructing 16 human organ and system aging clocks over 53,000 UK Biobank participants, this study reported over 5000 disease associations and over 6000 environmental associations, identified 261 genomic loci, and applied a Granger-like Bayesian regression to about 1000 longitudinal participants to infer potential temporal interaction networks. The findings highlight both organ-specific and pleiotropic genetic mechanisms underlying cross-organ aging interactions. This atlas serves as a unique resource for delineating genetic and environmental factors and the interaction network of multiple plasma protein-based organ and system aging, and may help inform the development of precise, targeted anti-aging strategies.

Methods

Study populations

UKB is a national cohort study that recruited approximately 500,000 participants aged 40–69 years old across the UK between 2006 and 2010. Baseline questionnaires and anthropometric measurements were collected from participants at 22 assessment centers throughout England, Wales, and Scotland. The full UKB protocol is available online (https://www.ukbiobank.ac.uk/media/gnkeyh2q/study-rationale.pdf). For the present analyses, we restricted our sample to UKB participants who had Olink Explore proteomic data available from baseline and who were drawn from the randomly selected subsample of the main cohort (n = 53,014).

The CKB is a prospective cohort study of 512,724 adults aged 30–79 years who were recruited from ten geographically diverse (five rural and five urban) areas across China between 2004 and 2008. The study design and methodological details have been described in prior publications⁴⁸. Our CKB analytic sample was limited to 3977 participants with baseline Olink Explore data available within a nested case–cohort study of ischemic heart disease (IHD), after excluding genetically related participants.

UKB has approval from the North West Multi-Center Research Ethics Committee as a Research Tissue Bank. CKB was approved by the Ethical Review Committee of the Chinese Center for Disease Control and Prevention, Beijing, China, and the Oxford Tropical Research Ethics Committee, University of Oxford. All participants consented to the use of their anonymized information for research purposes at the time of recruitment. All participants from the UKB and CKB provided written informed consent.

Proteomic profiling

Plasma proteomic profiling in both cohorts was conducted using the Olink Explore 3072 platform, which integrates four complementary panels: Cardiometabolic, Inflammation, Neurology, and Oncology. Preprocessed protein data across all cohorts were expressed as Normalized Protein eXpression (NPX) values on a log₂ scale. Within the UKB, the randomly selected proteomics subsample (n = 53,014) was derived by excluding participants from batches 0 and 7 collected in 2010. This random subsample has been previously demonstrated to be broadly representative of the overall UKB population⁴⁹, with full details on sample selection, processing, and quality control available through the UKB data portal. An additional 1463 proteins were quantified in plasma samples obtained at the imaging visit (2014 onwards) and at the first repeat imaging visit (2019 onwards) from 1172 and 1123 participants, respectively. Among these, 1006 participants with complete data across all three time points were included in longitudinal analyses.

In the CKB, baseline plasma samples were retrieved from long-term storage, thawed, and subdivided into multiple aliquots, with one 100 µl aliquot used to prepare two sets of 96-well plates (40 µl per well). The two plate sets were dispatched on dry ice to separate Olink facilities—one to the Olink Bioscience Laboratory in Uppsala (batch one; 1463 unique proteins) and the other to the Olink Laboratory in Boston (batch two; 1460 unique proteins)—with each batch covering the full sample of 3977 participants. Samples were plated in their retrieval order from the Wolfson Laboratory in Oxford. Normalization was performed using both an internal extension control and an inter-plate control, followed by transformation using a pre-specified correction factor. The limit of detection (LOD) was established from negative control samples (antigen-free buffer). Quality control warnings were assigned to samples in which the incubation control deviated by more than a predefined threshold (±0.3) from the median of all samples on the same plate; values falling below the LOD were nonetheless retained in the analyses.

Identification of organ-enriched plasma proteins

To identify organ-enriched genes and plasma proteins, we utilized the Gene Tissue Expression (GTEx) human tissue bulk RNA-seq database. Gene expression data across tissues were normalized using the DESeg2 R package. Organ-enriched genes were defined according to the Human Protein Atlas criteria: a gene was considered enriched if its expression level was at least four times higher in a single organ or system compared to any others⁵. Tissues from the same organ or system were grouped within GTEx, and the maximum expression value among the organ’s sub-tissues was designated as the representational expression level for that organ. For example, blood and spleen were defined as sub-tissues of the immune system. Enriched genes were then mapped to the 2923 plasma proteins quantified on the Olink Explore 3072 platform in the UKB.

Aging models and age gap calculation

Based on organ- and system-enriched plasma proteins (Supplementary Data 2), XGBoost (https://xgboost.readthedocs.io) was utilized to estimate organ age. XGBoost can automatically handle missing data by setting a default direction for missing values at each split node and learning this direction during training to maximize the gain of the split. For each organ and system, participants without any mapped organ and system-enriched proteins were excluded from 53,014 participants at baseline and the remaining were divided into the training set (70%) and the testing set (30%). Models were constructed to predict chronological age with sex as a covariate. Models were optimized within the training set through a fivefold cross-validation framework, employing a randomized hyperparameter search guided by Bayesian optimization (50 iterations). The hyperparameters explored were number of boosting rounds [seq(500:1500)], learning rate [seq(0.001:0.1)], maximum depth [seq(3:5)], gamma regularization parameter [seq(0:0.5)], column sample rate by tree [seq(0.6:1)], subsample rate [seq(0.5:1)], and min child weight [seq(1:5, by = 1)]. The remaining parameters were kept at default settings. Using the entire training dataset and the optimal hyperparameters, we recomputed aging models and applied them to predict organ age in the testing set. Model performance was quantified using the Pearson correlation coefficient (r) and MAE between predicted age and chronological age in the testing set. Shapley values were calculated to quantify the relative contribution of each organ- and system- enriched protein to its corresponding aging mode⁵⁰.

A linear regression model was then fitted between the predicted age and chronological age in each organ system, controlling for sex as a covariate. The age gap was calculated as the difference between the predicted age and the population means estimated by the regression model. To account for differences in model variability, age gaps were z−scored within each aging model. This allowed for direct comparison of age gaps across organ systems in downstream analyses (Supplementary Data 3).

Proteomic organ ages in CKB were predicted by directly applying the UKB-trained models to the CKB proteomic dataset. Predictive performance was quantified as the Pearson correlation (r) between predicted organ age and chronological age. To evaluate the validity of the transferred clocks, associations between organ age acceleration and all-cause mortality were examined using Cox proportional hazards models adjusted for age and sex.

Disease definition

In UKB, diagnoses and medical conditions of the participants were obtained through the linked hospital inpatient record data, self-reported data, and primary care data from the UK National Health Service. Diseases were identified using International Classification of Diseases-10 codes (ICD-10). Prevalent and incident diseases were classified as events that occurred before and after the date of the individuals’ baseline visit when the blood samples were collected. For incident diseases, follow-up time was calculated from the date of baseline assessment to the date of death, loss to follow-up, or end of follow-up (Dec 31, 2022), whichever came first. FinnGen disease endpoints code (https://www.finngen.fi/) was used to classify diseases, and related guidelines for quality control were applied, including pre-defined conditions (i.e., sex and age) and control exclusions (Supplementary Data 6). Incident diseases with fewer than 100 cases were excluded from analyses. A total of 655 incident and 404 prevalent disease endpoints were finally included. Delivery-related events (e.g., single spontaneous delivery and complications of labor and delivery) and certain gynecological diseases (e.g., menorrhagia and irregular menses) were excluded in further analysis, as they usually occur before 45 years old.

In CKB, incident diseases and all-cause mortality during follow-up were ascertained through data linkage of study participants to official death and disease registries and to national health insurance databases through December 31, 2018⁴⁸. Causes of death were obtained from official death certificates and, where necessary, supplemented by review of medical records. All diagnoses were coded using the ICD-10. We used the PheWAS R package⁵¹ to convert 1272 unique ICD-10 codes to a total of 633 hierarchical PheWAS codes. After excluding outcomes with fewer than 20 cases to ensure adequate statistical power, 109 incident diseases were retained for the analysis.

Environmental factors

UKB collected multidimensional environmental phenotypes on the baseline visit, such as lifestyles, environmental exposures, and physical measurements. PEACOK R package (https://github.com/astrazeneca-cgr-publications/PEACOK) was utilized to categorize the phenotypes into different data types. We further transformed ordered categorical variables that have a clear direction to health into continuous variables (e.g., neuroticism score). Phenotypes with less than 10,000 samples were excluded, resulting in 464 continuous phenotypes, 256 ordered categorical phenotypes, and 121 binary phenotypes (Supplementary Data 22). Continuous variables not following a normal distribution were log-transformed, and all continuous variables were standardized using z-score normalization to eliminate differences in scales for subsequent regression analysis.

Statistical analyses

Associations of multiple organ and system aging with diseases

Logistic regression models were performed for prevalent diseases, adjusting for chronological age and sex. Cox proportional hazards models with the same covariate adjustments were performed for incident diseases. Associations were considered statistically significant if they survived a stringent Bonferroni correction for multiple testing (prevalent diseases: P < 0.05/[16 × 404] = 7.74 × 10⁻⁶; incident diseases: P < 0.05/[16 × 655] = 4.77 × 10⁻⁶). Participants developing diseases within the first 2-year period were further excluded to attenuate potential reverse causation in the sensitivity analysis.

Associations of environmental factors with multiple organ and system aging

Diverse regression models were performed according to the variable type of environmental factors. Linear regression models were performed for the analysis of continuous factors. Logistic regression models were performed for binary factors. Proportional odds logistic regression models were performed for ordered categorical factors. All regression models were adjusted for the covariates mentioned above. Significance was determined using a stringent Bonferroni correction (P < 0.05/[16 × 841] = 3.72 × 10⁻⁶).

Genetic analyses

Genotyping and quality control

The UKB provides ~92 million variants across 487,442 participants, including imputed variants based on reference panels from the 1000 Genomes and the Haplotype Reference Consortium (HRC). We implemented a comprehensive quality control pipeline for both participant and variant filtering. For participant quality control, we retained individuals of European descent, identified through a detailed genetic principal components analysis. Participants were excluded if they exhibited discrepancies between genetic and self-reported sex or had sex chromosome aneuploidy. For variant quality control, only SNPs meeting the following criteria were retained: minor allele frequency (MAF) greater than 0.1%, an imputation quality score exceeding 0.3, and adherence to Hardy-Weinberg equilibrium with a P value greater than 1 × 10⁻⁶.

Genome-wide association analysis

We conducted GWAS on 16 organs and systems aging clocks to investigate the underlying genetic architecture. The fastGWA-MLM method was utilized through the Genome-wide Complex Trait Analysis software (version 1.94.1). Our models included chronological age, sex, genotype array, and the top 10 principal components as covariates. SNP P values reaching a threshold of P < 5 × 10⁻⁸ were considered statistically significant. To estimate SNP-based heritability (the variance explained by all SNPs), we used the genome-based restricted maximum likelihood approach within GCTA, incorporating the same set of covariates.

Functional mapping and annotation

We used the FUMA online platform (v1.5.2) for functional mapping and annotation of GWAS results, based on the human genome assembly GRCh37 (hg19) with default parameters. To identify independent genomic risk loci (r² > 0.6) and variants in linkage disequilibrium (LD) with lead SNPs, we utilized the SNP2Gene module, referencing genetic data from the European population in 1000 G phase 3. Positional mapping was performed with a 10 kb window to associate risk loci with nearby protein-coding genes.

Genetic correlation

We estimated pairwise genetic correlations between each pair of organs and systems using the linkage disequilibrium score regression (LDSC) software, utilizing precomputed linkage disequilibrium scores from the 1000 Genomes Project for individuals of European ancestry.

Gene-based association, gene-set, and gene-property analyses with MAGMA

Pertaining to genes and tissue enrichment analyses (i.e., GWGAS, gene-set, and tissue expression analyses) were performed using MAGMA v1.08 within FUMA. For each of the 18,955 protein-coding genes, GWGAS assessed the joint effect of all variants within the gene. Genome-wide significance was determined using Bonferroni correction (P < 0.05/18,955 = 2.64 × 10⁻⁶). To explore potential biological pathways or cellular functions linked to aging across organ systems, gene-set analysis was performed using hypergeometric tests on curated gene sets and Gene Ontology terms from MsigDB v7.0, applying Bonferroni correction for multiple testing.

Partitioned heritability estimate

To understand the contributions of distinct genomic categories to the heritability of aging across organs and systems, we employed partitioned heritability analysis via stratified LDSC. This approach quantifies the enrichment of heritability within predefined and annotated genome regions. We considered three sets of functional categories and cell- or tissue-specific types:

1) General Functional Categories

Partitioned heritability was calculated for 53 general functional categories, including coding, untranslated regions, promoters, introns, etc., without cell-type specificity⁵².

2) Cell-Type-Specific Categories

We estimated cell-type-specific partitioned heritability using gene sets from Cahoy et al.⁵³, focusing on three main cell types: astrocytes, neurons, and oligodendrocytes.

3) Tissue-Specific Categories

We used multi-tissue chromatin state data from ROADMAP⁵⁴ and ENCODE⁵⁵, as well as multi-tissue gene expression data from GTEx V8⁵⁶, to assess tissue-specific partitioned heritability. Bonferroni correction was applied to all tested annotations and categories. Linkage disequilibrium scores and allele frequencies for European ancestry were derived from the 1000 Genomes Project. The methodologies for stratified LDSC are detailed in the original work⁵⁷.

Longitudinal analyses of organ and systems aging

Re-trained longitudinal organ age

Based on the repeated plasma proteomics data from UKB, we re-trained organ age models among 1006 participants with available data on 1463 proteins that were measured in three visits. To account for within-subject correlations in repeated-measures, we employed the GroupShuffleSplit function from sklearn.model_selection Python package during model training. This approach ensures that all follow-up data from the same individual are assigned to the same fold. The remaining model hyperparameters were kept consistent with those of the baseline model.

Rates of aging in each organ and system were calculated respectively among the 1006 participants using generalized linear mixed models (GLMM), adjusting for chronological age at baseline and sex:

$$\mathrm{Organ}{\mathrm{Age}}_{\mathrm{ij}}={\mathrm{\beta }}^0+{\mathrm{x}}_{\mathrm{ij}}^{\mathrm{\Delta }\mathrm{age}}{\mathrm{\beta }}^{\mathrm{slope}}+{\mathrm{x}}_{\mathrm{ij}}^{\mathrm{\Delta }\mathrm{age}}{\mathrm{u}}_{\mathrm{i}}^{\mathrm{slope}}+{\mathrm{\epsilon }}_{\mathrm{ij}}$$ 1

Organ age_ij refers to the predicted organ age for the i th participant on the j th visit, ${\beta }^0$ is a fixed intercept, $x_{\mathrm{i}j}^{\mathrm{\Delta }\mathrm{age}}$ is the difference in chronological age between the j th visit and the age at baseline for the i th participant. ${\beta }^{\mathrm{slope}}$ is the fixed effect index indicating the overall change in predicted organ age per year of chronological age, while ${\mathrm{u}}_{\mathrm{i}}^{\mathrm{slope}}$ is the random effect indicating that within every participant. ${\mathrm{\epsilon }}_{\mathrm{ij}}$ is a random error term. GLMM was performed by the lme4 R package. The rate of organ aging of the ith participant was calculated as ${\beta }^{\mathrm{slope}}$ + ${\mathrm{u}}_{\mathrm{i}}^{\mathrm{slope}}$. T-tests were utilized to examine the influence of sex and baseline age gaps on organ and system aging rates.

Multitask Granger-like Bayesian regression framework with automatic relevance determination

To investigate the dynamic temporal interactions underlying aging across organs and systems, we employed a multitask Granger-like Bayesian regression framework with ARD to construct a directed network of aging-related interdependencies among organs and systems. Longitudinal data comprising organ system-specific biological age gaps, defined as deviations from chronological age, were analyzed across multiple time points. For each organ $j$, its aging metric at time $t+1,Y_j^{t+1}$ was modeled as a function of the aging states of all other organs at the preceding time point$t{X}_j^t$, adjusted for covariates including chronological age, gender, and longitudinal study time. This temporal design explicitly encodes the principle of causal precedence, wherein putative causal drivers $X^t$ temporally precede their effects $Y^{t+1}$, aligning with Granger causality assumptions.

The modeling framework treated the prediction of each organ’s future state as a distinct task within a multitask learning paradigm. For target organ j, the regression model took the form:

$$Y_j^{t+1}={\beta }_{0j}+{\sum }_{i\ne j}{\beta }_{ij}{X}_i^t+{\gamma }_{1j}Age+{\gamma }_{2j}Gender+{\gamma }_{3j}TimeGap+{\epsilon }_j$$ 2

Where ${\beta }_{ij}$ quantified the putative causal influence of organ $i$’s aging on organ $j$, while ${\gamma }_{1j}$, ${\gamma }_{2j}$ and ${\gamma }_{3j}$ represented the effects of the covariates: chronological age (continuous), gender (binary, encoded as 0 for male and 1 for female), and longitudinal study time (time elapsed since baseline). To enforce sparsity and mitigate overfitting, hierarchical Bayesian regularization was implemented through ARD. Each regression coefficient ${\beta }_{ij}$ was assigned a zero-mean Gaussian prior with a precision hyperparameter ${\alpha }_{ij}$:

$${\beta }_{ij}~N\left(0,{\alpha }_{ij}^{-1}\right)$$ 3

The hyperparameters ${\alpha }_{ij}$ were inferred from the data, automatically shrinking coefficients for non-informative predictors toward zero, thereby pruning spurious connections. Crucially, hyperparameters were shared across tasks, enabling the identification of globally irrelevant predictors (e.g., organs exerting no causal influence on most targets) and enhancing the robustness of network inference.

The entire analysis workflow was implemented in the R environment. Data management and statistical computations were performed using base R packages and dplyr, including data structure design, covariate adjustment, and result visualization. The modeling process leveraged the reticulate package to call and utilized the sklearn.linear_model.ARDRegression module to implement the core algorithm for Bayesian sparse regression.

The final temporal network retained only connections with non-negligible regression coefficients ${\beta }_{ij}$ after ARD-driven sparsity constraints, prioritizing parsimony and interpretability. This approach integrated temporal precedence as a foundational causal assumption, Bayesian regularization for feature selection, and multitask learning to map directional interactions in organ aging. The resulting network highlighted organ system-specific drivers of aging dynamics while controlling for systemic confounders such as age-related trends and sex differences.

Reporting summary

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

Author contributions

Z.Y.L. designed the study. X.Q.J. and Y.H.W. conducted the data analysis. F.P., X.Q.J., X.X.H., and C.Y.Z. drafted the manuscript. Y.J.Z., J.N.Y., S.M.W., S.H.Wu (Shuhui Wu), S.H.Wang (Shenghan Wang), D.D.F., D.Q.C., Y.S.W., Z.Y.Z., D.W., N.L., J.F.C., X.F.W., J.J., C.H., A.A.C., H.D.D., D.B., R.W., C.Q.Y., J.L., Z.M.C., L.M.L., L.J., D.S., Z.X.H., and Z.Y.L. critically revised the manuscript for important intellectual content. All authors approved the final version of the manuscript. Z.Y.L. is the guarantor. The corresponding author attests that all the listed authors meet the authorship criteria and that no others meeting the criteria have been omitted.

Peer review

Peer review information

Nature Communications thanks Raghav Sehgal and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This research was supported by grants from the National Natural Science Foundation of China (824B1013, 72374180, 82103920), “Pioneer” and “Leading Goose” R&D Programs of Zhejiang Province (2026C02A1147, 2025C02104), National Key R&D Program (2023YFC2509400), and the Shanghai Municipal Science and Technology Major Project630631(grants 2023SHZDZX02 and 2017SHZDZX01 to L.J.). The CKB baseline survey and the first re-survey were supported by the Kadoorie Charitable Foundation in Hong Kong. The long-term follow-up has been supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2023ZD0510101, 2023ZD0510100), the National Natural Science Foundation of China (82192900, 81390540, 82388102, 91846303, 81941018), the National Key R&D Program of China (2016YFC0900500), the Chinese Ministry of Science and Technology (2011BAI09B01), and Wellcome grants to Oxford University (212946/Z/18/Z, 202922/Z/16/Z, 104085/Z/14/Z, 088158/Z/09/Z). The UK Medical Research Council (MC_UU_00017/1, MC_UU_12026/2 and MC_U137686851), Cancer Research UK (C16077/A29186 and C500/A16896) and the British Heart Foundation (CH/1996001/9454) provide core funding to the Clinical Trial Service Unit and Epidemiological Studies Unit at Oxford University for the project. The CKB proteomic assays were supported by the British Heart Foundation (18/23/33512), Novo Nordisk and OLINK. This research was also supported by grants from Hangzhou Meilian Medical Co., Ltd (Kheng-20241116, to ZL), Zhejiang Key Laboratory of Intelligent Preventive Medicine (2020E10004), Zhejiang University Global Partnership Fund, and Zhejiang University School of Public Health Interdisciplinary Research Innovation Team Development Project. The funders had no role in the study design; data collection, analysis, or interpretation; in the writing of the report; or in the decision to submit the article for publication.

Data availability

UKB data are available through a procedure described at https://www.ukbiobank.ac.uk/enable-your-research. The present analysis of UK Biobank data was conducted under UK Biobank application number 61856. Access can be obtained by bona fide researchers through an application to UKB. The CKB is a global resource for the investigation of lifestyle, environmental, blood biochemical and genetic factors as determinants of common diseases. The CKB study group is committed to making the cohort data available to the scientific community in China, the UK and worldwide to advance knowledge about the causes, prevention and treatment of disease. For detailed information on what data are currently available to open-access users, how to apply for them and the timeline for data access, visit the CKB website at https://www.ckbiobank.org/data-access. Researchers who are interested in obtaining the raw data that have been officially released from the CKB study should contact ckbaccess@ndph.ox.ac.uk. A research proposal will be requested to ensure that any analysis is performed by bona fide researchers. For any data that is not currently available via open access, researchers may need to develop a formal collaboration with the CKB study group. Source data are provided with this paper.

Code availability

All analyses were performed in R version 4.2.1 (http://www.R-project.org) and publicly available packages. The survival analyses is performed using the R package survival (version 3.8.3, https://github.com/therneau/survival). XGBoost was performed using the R package xgboost, with hyperparameter tuning performed using the R packages mlr (version 2.19.2, https://github.com/mlr-org/mlr) and mlrMBO (version 1.1.5.1, https://github.com/mlr-org/mlrMBO). Other codes used for data processing and statistical analyses are available from the corresponding author upon reasonable.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77365-8.