Abstract
Animal studies show aging varies between individuals as well as between organs within an individual1,2,3,4, but whether this is true in humans and its effect on age-related diseases is unknown. We utilized levels of human blood plasma proteins originating from specific organs to measure organ-specific aging differences in living individuals. Using machine learning models, we analysed aging in 11 major organs and estimated organ age reproducibly in five independent cohorts encompassing 5,676 adults across the human lifespan. We discovered nearly 20% of the population show strongly accelerated age in one organ and 1.7% are multi-organ agers. Accelerated organ aging confers 20–50% higher mortality risk, and organ-specific diseases relate to faster aging of those organs. We find individuals with accelerated heart aging have a 250% increased heart failure risk and accelerated brain and vascular aging predict Alzheimer’s disease (AD) progression independently from and as strongly as plasma pTau-181 (ref. 5), the current best blood-based biomarker for AD. Our models link vascular calcification, extracellular matrix alterations and synaptic protein shedding to early cognitive decline. We introduce a simple and interpretable method to study organ aging using plasma proteomics data, predicting diseases and aging effects.
Main
Aging results in organism-wide deterioration of tissue structure and function that drastically increases the risk of most chronic diseases. Comprehensive studies of the molecular changes that occur with aging across multiple organs in mice have identified unique molecular aging trajectories and timings1,2,3,4, and susceptibility and resilience to diseases of aging in specific organs such as the brain, heart and kidney varies substantially across the population6. However, little is known about how human organs change molecularly with age. A molecular understanding of human organ aging is of critical importance to address the massive global disease burden of aging and could revolutionize patient care, preventative medicine and drug development7. In particular, preclinical studies have demonstrated that rejuvenating interventions affect organs differently3,8. To translate these studies into transformative medicines, we must be able to accurately measure aging across the body and understand the diversity of human aging not only across but also within individuals.
While many methods to measure molecular aging in humans have been developed9,10,11, most of them provide just a single measure of aging for the whole body. This is difficult to interpret given the complexity of human aging trajectories. Some recent methods have used clinical chemistry markers which include some markers of organ function12,13,14,15. However, many of these markers have low organ specificity, making them difficult to interpret for organ-specific aging. Methods to measure brain aging have used MRI-based brain volume and functional connectivity measurements, which are costly and do not provide molecular insights16, or have required tissue samples, which prevents their application in living persons17. Building off the wealth of literature and clinical practice that uses certain organ-specific plasma proteins to noninvasively assess aspects of organ health, such as alanine transaminase for liver damage, we hypothesized that comprehensive quantification of organ-specific proteins in plasma could enable minimally invasive assessment and tracking of human aging for any organ.
Plasma proteins can model organ aging
To test this, we measured 4,979 proteins in a total of 5,676 subjects across five independent cohorts (Supplementary Table 1) and mapped the putative organ-specific plasma proteome, which we used to train models of organ aging (Fig. 1a). We mapped the organ-specific plasma proteome using human organ bulk RNA sequencing (RNA-seq) data from the Genotype-Tissue Expression (GTEx) project18. We classified genes as ‘organ enriched’ if they were expressed at least four times higher in one organ compared to any other organ, according to the definition proposed in the Human Protein Atlas19 (Extended Data Fig. 1, Supplementary Tables 2 and 3, and Methods). We annotated the 4,979 human proteins measured by the SomaScan assay with this information and found 893 (18%) proteins met this definition, with the highest number from the brain. We performed additional quality control to remove proteins with a high coefficient of variation or a low correlation between the two different versions of the SomaScan assay present across our cohorts, leaving us with 4,778 proteins (856 organ enriched, 17.9%) which were used for downstream analysis (Supplementary Fig. 1 and Supplementary Tables 4 and 5).
We and others have previously shown that plasma proteins can be used to train machine learning models to estimate chronological age in independent cohorts20,21. For each individual, an aging model produces an ‘age gap’, a measure of that individual’s biological age relative to other same-aged peers based on their molecular profile9 (Fig. 1a). Several studies have shown associations between age gaps and mortality risk or other age-related phenotypes9, supporting the hypothesis that the age gap contains information about relative biological aging.
Based on this concept, we trained a bagged ensemble of least absolute shrinkage and selection operator (LASSO) aging models for 11 major organs using the mutually exclusive organ-enriched proteins we identified as inputs (Fig. 1a, Extended Data Fig. 2a,b, Supplementary Fig. 3 and Supplementary Tables 6–8). We chose to restrict our analyses to adipose tissue, artery, brain, heart, immune tissue, intestine, kidney, liver, lung, muscle and pancreas because of their relatively well-understood contributions to diseases of aging and the availability of relevant age-related phenotype data in the tested cohorts. We also trained an ‘organismal’ aging model using the 3,907 organ-nonspecific plasma proteins as inputs to compare the contribution of specific organs to an organ-shared aging signature, and a ‘conventional’ proteomic aging model using all 4,778 proteins to compare the organ aging models to a global plasma proteomic aging signature as previously reported20,21. We trained our models in 1,398 healthy participants from the Knight Alzheimer’s Disease Research Center (Knight-ADRC) cohort (mean age = 75, age range = 27–104) and then tested these models in four fully independent cohorts and in held-out test participants with dementia in the Knight-ADRC. (Fig. 1a, Extended Data Figs. 2 and 3, and Supplementary Fig. 2). All 11 organ aging models and the organismal model significantly estimated age in all five cohorts after multiple test correction (Supplementary Fig. 3b). Organ-specific proteins selected by our approach were highly enriched for organ-specific functions (Supplementary Information).
We observed across all cohorts that individuals with the same conventional age gap had diverse organ aging profiles (Fig. 1b). At the population level, this resulted in a low-to-moderate correlation between the age gaps of different organs (mean pairwise Pearson r = 0.29, Fig. 1c). While organ aging is correlated, the majority of variance in one organ age gap is not explained by others, with the exception of the organismal and conventional age gaps which were highly correlated. Further, we observed that some individuals had extreme aging in one or more organs relative to the general population (Fig. 1d). We scored individuals across all cohorts as outliers for a given organ age gap using a two standard deviation cutoff and clustered individuals into extreme aging types (e-ageotypes) (Fig. 1e and Extended Data Fig. 4a–c). Although it might be expected that extreme aging in one organ would co-occur with extreme aging in other organs, we instead observed segregation into distinct organ e-ageotypes. We found that approximately 18.4% of individuals had a highly organ-specific e-ageotype that was dominated by the aging of only one organ. Only approximately 1.7% of individuals showed extreme aging in multiple organs; the only multi-organ e-ageotype discovered through unbiased clustering was defined by extreme adipose, brain, conventional, heart, immune, liver and organismal age gaps. These observations suggest that organ age gaps may capture unique aging information, which may have implications for organ-specific biological aging and diseases of aging.
Organ age predicts health and disease
To assess the relationship between organ age and biological aging, we tested whether organ e-ageotypes were associated with nine age-related disease states for which we had sufficient data in at least two independent cohorts; AD, atrial fibrillation, cerebrovascular disease, diabetes, heart attack, hypercholesterolaemia, hypertension, obesity and gait impairment. Organ e-ageotypes were associated with specific disease states with known high impact on their respective organs (23 of 117, 20%, associations significant in a meta-analysis after multiple testing correction, Extended Data Fig. 4d and Supplementary Table 9). The kidney ageotype was the most significantly associated with metabolic diseases (diabetes, obesity, hypercholesterolaemia and hypertension), the heart ageotype was the most significantly associated with heart diseases (atrial fibrillation and heart attack), the muscle ageotype was the most significantly associated with gait impairment, the brain ageotype was the most significantly associated with cerebrovascular disease and the organismal ageotype was the most significantly associated with AD. At the whole population level, the relationships between organ age gaps and disease showed the same trends as ageotypes, but more diseases were significantly associated with age gaps due to higher statistical power (65 of 117, 56%, statistically significant after multiple test correction, Extended Data Fig. 4e and Supplementary Table 10).
At the population level, the two most significant associations between disease and age gap were between the kidney age gap and metabolic disease traits. Individuals with hypertension had kidneys that were approximately one year older than their same-aged peers, while individuals with diabetes had kidneys approximately 1.3 years older (Fig. 2a,b and Supplementary Tables 8 and 10). The third and fourth top associations were between the heart age gap and the heart aging traits atrial fibrillation (2.8 years older) and heart attack (2.6 years older) (Fig. 2c,d). Overall, we found that certain diseases, such as heart attack and AD, were associated with accelerated aging in virtually all organs, while others had impacts on a particular organ or subset of organs (Extended Data Fig. 4e and Supplementary Table 10).
Kidney aging proteins were highly expressed by kidney cell types (Fig. 2e,f) and had known roles in kidney biology and disease. Using feature importance plots, the model identified renin (REN), a kidney enzyme known to regulate blood pressure via the renin-angiotensin pathway22, as an important protein in kidney aging. It also identified the putative longevity factor klotho (KL)23, as well as multiple proteins with unknown functions including uromodulin (UMOD) and kidney associated antigen 1 (KAAG1), as important kidney aging proteins. UMOD has been genetically linked to chronic kidney disease, where it is observed to have age-dependent effects24, and rare mutations are the major cause of autosomal dominant tubulointerstitial kidney disease25.
Heart aging proteins were expressed primarily by cardiomyocytes (Fig. 2g,h) and had known roles in heart biology and disease. Pro-brain natriuretic peptide (NPPB), a negative regulator of blood pressure that increases in response to heart damage, and troponin T (TNNT2), a heart muscle protein involved in contraction, had the strongest weights in the heart aging model (Fig. 2g). They are both established clinical markers of acute heart failure26, and NPPB has been previously associated with heart attack risk27. This suggests the possibility of a link between subclinical heart disease and the ‘normal’ heart aging process, which should be investigated further with more detailed heart imaging and electrophysiology. Less well-characterized heart proteins include cardiac myosin light chain (MYL7), peroxidasin like (PXDNL) and bone morphogenetic protein 10 (BMP10). MYL7 is expressed by atrial cardiomyocytes and has recently become a promising target for hypertrophic cardiomyopathy28, suggesting that this could be a repurposing target for heart aging more generally.
Given the strong associations between heart aging traits and the heart age gap, we used longitudinal follow-up among healthy participants in the LonGenity cohort to test if organ age was significantly associated with future heart failure risk (Fig. 2i and Supplementary Table 11). We found that among people with no active disease or clinically abnormal biomarkers at baseline, every 4.1 years of additional heart age (one standard deviation) conferred an almost 2.5-fold increased risk of heart failure over a 15-year follow-up (23% increased risk per year of heart aging, Fig. 2i). Age gaps from multiple other tissues, but not the conventional aging model, also trended towards significance.
We next tested the associations between organ age gaps and all-cause mortality. We found that the age gaps from 10 out of 11 organs, the organismal model and the conventional model were significantly associated with future risk of all-cause mortality after multiple test correction in the LonGenity cohort over 15 years of follow-up (Fig. 2j and Supplementary Table 12). A standard deviation increase (approximately four years of extra organ aging, Supplementary Table 8) in heart, adipose, liver, pancreas, brain, lung, immune or muscle age gap each conferred between 15–50% increased all-cause mortality risk. These hazard ratios are a similar size to methylation-based mortality predictors in independent aging cohorts over similar follow-up times, despite the fact that organ aging models are trained to predict chronological age instead of mortality directly (DNAm GrimAge hazard ratio = 1.3, 14 year mortality follow-up29). Further, we found that for some organs, there was a nonlinear relationship between the age gap and mortality risk (Supplementary Information, Supplementary Fig. 4 and Supplementary Table 13).
Finally, to better understand the relationship between organ age and additional markers of health and disease, we tested the associations between organ age gaps and 43 clinical biochemistry and cell count markers in the test cohort Covance (Extended Data Fig. 5 and Supplementary Fig. 5, see Supplement Information for additional discussion). We also used these markers to calculate Phenotypic age14 (PhenoAge), a clinical biochemistry-based aging clock which predicts mortality and morbidity risk, for all participants in Covance (Extended Data Fig. 5a). We found that the PhenoAge age gap was significantly correlated with multiple organ age gaps, but only a small portion of the variance in any model was explained by another (Extended Data Fig. 5b).
We found 226 out of 559 (40%) associations between organ age gaps and clinical biochemistry markers were significant after multiple testing correction (Extended Data Fig. 5c and Supplementary Table 14). The strongest associations included associations between liver age gap and blood AST:ALT ratio, a clinical marker of liver health and function that is known to change with age (adjusted Pearson r = 0.25, q = 6.13 × 10−17), and between kidney age gap and serum creatinine, the standard clinical marker of kidney function (adjusted Pearson r = 0.23, q = 1.65 × 10−16). While these results are highly significant, they only partially explain the relationship between organ age gaps and disease phenotypes. Even after correcting for estimated glomerular filtration rate (eGFR), the kidney age gap is still significantly associated with hypertension and diabetes (Supplementary Fig. 6).
Collectively, organ age gap associations with disease and blood biochemistry demonstrate that aging models derived from organ-specific plasma proteins capture disease-relevant heterogeneity of aging within and across individuals, which is not captured by other aging clocks or clinical markers.
Brain aging in cognitive decline and AD
Although the largest risk factor for neurodegenerative diseases is age, little is known about the contribution of molecular brain aging to disease. The brain age gap correlated significantly with AD in held-out participants in the Knight-ADRC, but did not replicate in the Stanford Alzheimer’s Disease Research Center (Stanford-ADRC) (Supplementary Table 10). Therefore, to better understand how underlying proteins contributed to the brain aging model’s predictive abilities for brain aging phenotypes, we developed the feature importance for biological aging (FIBA) algorithm, which uses feature permutation to generate a per-protein importance score for both chronological and biological age, as defined by a particular age-related trait (Extended Data Fig. 6a and Methods). We applied FIBA to the brain age model using the trait global clinical dementia rating (CDRGLOB) in the Knight-ADRC cohort to understand how brain proteins contributed to the association between the age gap and cognitive decline. We observed that some proteins, such as complexins, increased both the model age prediction accuracy and the age gap association with dementia severity (FIBA+), while others decreased the age gap association with dementia severity (FIBA−) (Fig. 3a and Supplementary Table 15).
We used this information to train a second-generation brain aging model, which we term the CognitionBrain aging model, by only using CDRGLOB FIBA+ brain-specific proteins (Fig. 3b and Supplementary Tables 16–19). This method is similar to second-generation methylation aging clocks which are trained jointly on chronological age and aging phenotypes14. We found that the CognitionBrain age gap had a stronger association with AD than the first-generation brain age gap and the conventional age gap in the Knight-ADRC cohort (Extended Data Fig. 6b). This result replicated in the independent test cohort Stanford-ADRC. In a meta-analysis, individuals with AD had approximately two years of additional CognitionBrain aging (P valuemeta = 9.23 × 10−36) compared to individuals without AD (Fig. 3c and Supplementary Table 20). The CognitionBrain age gap was also significantly associated with risk of future dementia progression in both ADRC cohorts. A standard deviation increase in the CognitionBrain age gap conferred a 34% increased risk (P valuemeta = 1.03 × 10−15) of a clinically relevant two-point increase in the Clinical Dementia Rating Sum-of-Boxes score (CDR-SB) within five years (Supplementary Table 21). We also tested associations between CognitionBrain age gap and changes in brain volume using matched volumetric MRI in the Stanford-ADRC and Stanford Aging and Memory Study (SAMS) cohorts (Extended Data Fig. 6c, Supplementary Table 22, Supplementary Fig. 7 and Supplementary Information), and found CognitionBrain age gap significantly predicted brain volume in multiple AD-sensitive regions.
Given its associations with AD status, cognitive decline risk and brain volume, we asked whether the CognitionBrain aging model could be used in combination with other biomarkers of AD and predictors of cognitive decline, including plasma pTau-181 (ref. 5) and an AD polygenic risk score30, to better stratify AD patients for future clinical outcomes. We tested a multivariate dementia progression cox proportional hazard model with baseline CDRGLOB, age, CognitionBrain age gap, plasma pTau-181 and an AD polygenic risk score (Fig. 3d) in the Stanford-ADRC. We found that the CognitionBrain age gap had the highest adjusted hazard ratio (hazard ratio = 1.57; P = 8.95 × 10−3) of the AD biomarkers, and that both plasma pTau-181 and CognitionBrain age gap were additive for risk prediction (estimated combined hazard ratio = 2.08, Fig. 3e). Individuals with fluid biomarker levels two standard deviations above average had a 75% probability of dementia progression, while individuals with levels two standard deviations below average had under a 10% probability of dementia progression within five years. Pairwise correlation between all biomarkers also showed that the CognitionBrain age gap was largely independent from other biomarkers (Extended Data Fig. 6d). Taken together, these data suggest CognitionBrain age gap provides molecular information about brain aging not captured by other approaches.
Given the significant associations between the CognitionBrain age model and several brain aging metrics, we sought to uncover new insights into brain aging mechanisms by examining the proteins that make up the model. A total of 47 of the 49 model proteins were detectable in human brain single-cell RNA sequencing (scRNA-seq) data and most could be mapped to neurons and glia with high specificity (Fig. 3f). Proteins with the largest positive weights in the model (Fig. 3c) included the synaptic proteins complexin 1 (CPLX1), complexin 2 (CPLX2) and neurexin 3 (NRXN3)—which all have genetic links to cognition and AD31,32,33—and stathmin 2 (STMN2) and olfactomedin 1 (OLFM1)—which are involved in neurite outgrowth and axon growth cone collapse34,35. Proteins with large negative weights in the model such as Aldolase Fructose-Bisphosphate C (ALDOC), neuronal pentraxin receptor (NPTXR), carnosine dipeptidase 1 (CNDP1) and Lanc Like Glutathione S-Transferase 1 (LANCL1). ALDOC, NPTXR and CNDP1 are expressed in astrocytes, neurons and oligodendrocytes, respectively (Fig. 3f) and have been proposed as CSF biomarkers for AD36,37. LANCL1, which is primarily expressed in oligodendrocytes (Fig. 3f), has been shown to be crucial for neuronal health in mouse models38. The model also implicated alterations in the glycosylated extracellular matrix through the proteins tenascin R (TNR), neurocan (NCAN) and heparan sulfate-glucosamine 3-sulfotransferase 4 (HS3ST4), underlining the role of the extracellular matrix in brain aging.
We assessed the highest weighted CognitionBrain proteins for their changes with age and AD in the Knight-ADRC and Stanford-ADRC cohorts, as well as their changes with AD in brain tissue at the protein39, bulk RNA39 and single-cell RNA levels from publicly available datasets (Fig. 3g). We observed a consistent pattern of decreases in AD brain tissue and increases in the blood with age and AD. This suggests that the increase of synapse and neurite growth related protein levels in the blood could reflect a loss or alteration in protein processing and subsequent shedding of these crucial factors in the brain. A similar inverse relationship between fluid and brain protein levels is seen with amyloid beta, whereby lower CSF AB42 is correlated with increased AB plaques in the brain40.
Organ aging in cognitive decline and AD
We next sought to apply the FIBA optimization framework to other organ aging models to understand how the aging of other organs contributes to brain aging phenotypes (Fig. 4a). As with the brain aging model, we applied CDRGLOB FIBA to all aging models using the Knight-ADRC (Extended Data Figs. 7 and 8). The CognitionArtery, CognitionBrain, CognitionOrganismal and CognitionPancreas age gap associations with AD replicated in both ADRCs (Fig. 4b and Extended Data Fig. 8c,d), so we focused on these four aging models to understand peripheral versus central contributions to cognitive decline.
To understand the full temporal sequence of cognitive decline, we tested if age gaps were associated with cognition in cognitively normal individuals using a composite score of overall cognition in the LonGenity cohort. The decreased cognitive function was significantly associated with all four age gaps (Fig. 4c, Extended Data Fig. 9a and Supplementary Table 23). We replicated these associations in the healthy SAMS cohort, where we observed that individuals with worse memory recall had higher CognitionOrganismal and CognitionBrain age gaps (Extended Data Fig. 9b and Supplementary Table 23).
We next tested associations between age gaps and risk of transition from cognitively normal to mild cognitive impairment (MCI) (CDR-Global Score 0 to greater than or equal to 0.5) using 15 years of clinical cognitive assessment in the Knight-ADRC (Fig. 4d and Supplementary Table 24). We found that the CognitionOrganismal (hazard ratio = 1.17, P = 0.02) and CognitionArtery (hazard ratio = 1.15, P = 0.04) age gaps significantly predicted conversion to MCI, while the CognitionBrain (hazard ratio = 1.11, P = 0.14) trended towards significance (Fig. 4d). The prediction of future conversion to MCI over 15 years is unlikely to be explained by undiagnosed cognitive impairment, placing changes detected by these aging models early in the causal chain of cognitive decline and neurodegenerative disease.
To understand the biological processes and proteins involved in early cognitive decline, we plotted the aging trajectory of all model proteins and found that highly weighted CognitionOrganismal and CognitionArtery proteins changed with age earlier and at a faster rate than CognitionBrain and CognitionPancreas proteins (Fig. 4e). The earliest changes occurred in a highly correlated cluster of CognitionOrganismal proteins: pleiotrophin (PTN), transgelin (TAGLN), WNT1 Inducible Signalling Pathway Protein 2 (WISP2), CUB Domain Containing Protein 1 (CDCP1) and chordin like 1 (CHRDL1; Fig. 4f). Though not organ-specific, these genes were all highly expressed in the arteries and brain (Extended Data Fig. 10a). Single-cell expression of these genes in human vasculature41,42, indicated these genes are expressed primarily by smooth muscle cells, pericytes and fibroblasts (Fig. 4g and Extended Data Fig. 10b). Loss of brain pericytes, smooth muscle cells and perivascular fibroblasts is associated with age and AD42,43 (Fig. 4g), and pericyte-specific deletion of PTN renders neurons prone to ischaemic and excitotoxic injury44. This early changing signature in the CognitionOrganismal model may thus represent degenerative changes to the cellular integrity of the brain vasculature and the loss of its neuroprotective functions with aging (Fig. 4h).
The five proteins composing the CognitionArtery model, TNF receptor superfamily member 11b (TNFRSF11B), sclerostin (SOST), melanocortin 2 receptor accessory protein (MRAP2), frizzled related protein (FRZB) and matrix gla protein (MGP) were also primarily expressed in vascular smooth muscle cells, pericytes and fibroblasts41 (Extended Data Fig. 10c) and are all strongly implicated in vascular calcification. TNFRSF11B/APOE double knockout mice show increased calcium deposition by vascular smooth muscle cells45, MGP deficiency-causing mutations in humans leads to Keutel syndrome, a disease characterized by soft tissue calcification46, and SOST and FRZB are negative regulators of WNT signalling that drive calcification and are increased in the plasma of people with vascular calcification47,48. We found that CognitionArtery proteins and the vascular signature in the CognitionOrganismal proteins form an interaction network using StringDB (Fig. 4i). Additional model proteins in this interaction network included integrin binding sialoprotein (IBSP), osteoglycin (OGN), collagen type III alpha 1 chain (COL3A1), proline rich and gla domain 1 (PRRG1) and growth arrest specific 6 (GAS6). In total, this protein network is involved in extracellular matrix, cartilage development and osteoblast signaling pathways, and implicates vascular calcification and extracellular matrix alterations as a major component of aging that underlies the early phases of cognitive decline and neurodegenerative disease.
Discussion
Our study introduces a framework for modelling organ health and biological aging using plasma proteomics. The resulting organ aging models can predict mortality, organ-specific functional decline, disease risk and progression and aging heterogeneity between tissues. This approach is minimally invasive, requiring only a small blood sample, and could be easily applied to understand the effects of health interventions, such as lifestyle modifications and drug therapies, at the organ level. We provide a large and comprehensive resource of organ aging information in nearly 6,000 individuals spanning the adult lifespan and multiple age-related disease states, and we have developed an easy-to-use python package called organage to calculate the organ ages of any plasma proteomics sample from the SomaScan assay.
There are many future directions for this work. While we have shown that plasma proteomic organ aging models are distinct from previous proteomics models, clinical chemistry-based models and imaging-based models, future studies should assess how proteomic organ aging relates to other molecular measures of aging and disease such as methylation aging clocks and disease-specific prediction models. Although we were unable to perform direct comparisons, our models predict mortality with comparable effect sizes to models trained specifically to predict mortality and heart disease in independent cohorts49,50. We demonstrated that our approach added increased value to established biomarkers of AD, and we expect that multimodal aging and disease prediction models may have similar impacts in other diseases.
We present one of the largest studies of plasma proteome aging to date, but as larger plasma proteomics resources emerge, the power of this approach will further increase. Our current models rely on approximately 5,000 proteins measured with the SomaScan assay, but the approach is platform agnostic, and we expect that even more biological information could be gained with additional proteomic coverage, including cell and organ-specific splice isoforms and posttranslational modifications. The rapidly growing number of human gene expression maps at single-cell resolution41 will help further refine organ and cell-type specific aging models and allow for a comprehensive understanding of organismal physiology based on the plasma proteome.
Another question for future studies is which organ-specific aging proteins are causal drivers of aging, given that multiple plasma proteins have been shown to directly modulate aging phenotypes8. Of note, many of the proteins with large weights in the models, such as KLOTHO, UMOD, MYL7, CPLX1, CPLX2 and NRXN3, have genetic associations with diseases of their respective organs or are validated therapeutic targets, suggesting a potential causal role of these proteins in organ aging. Future genomic studies should further investigate the genetic architecture of organ aging clocks and their relationships to disease using GWAS and post-GWAS methods such as colocalization and Mendelian randomization.
This study has multiple limitations. First, we have limited the study to a subset of organs to avoid over-interpretation of models for which we lacked convincing organ-relevant aging phenotypes. It remains unclear if this approach will generalize to all organs in the body, such as reproductive organs, and future studies should address this question. Second, we observe many instances of nonlinear dynamics in the plasma proteome and in aging phenotypes. While our current models serve as a proof of principle for this approach, since they are trained and evaluated largely on older adults, caution should be used when applying them to young people. More sophisticated nonlinear machine learning methods such as neural networks or random forests may further improve the accuracy and generalizability of this approach in the future. Lastly, the models were trained and tested on American and Caucasian-skewed cohorts, and future studies should assess the generalizability of the findings in more ethnically and geographically diverse populations.
Altogether, we show that large-scale plasma proteomics and machine learning can be leveraged to noninvasively measure organ health and aging in living people. We show that biologically motivated modelling, in which we use sets of organ-specific proteins and the FIBA algorithm to further subset to physiological age-related proteins, enables deconvolution of the different rates of aging within an individual and measurement of aging at organ-level resolution.
Methods
Human cohorts
Covance
Details of the Covance study have been previously published54. Briefly, Covance is a multi-site cross-sectional study of health across the lifespan collected at five hospital sites in the United States in 2008. A total of 1,028 subjects were included in analyses for this study. Cohort demographic characteristics are summarized in Supplementary Table 1. Exclusion criteria for the study included uncontrolled hypertension, self-reported treatment for a malignancy other than squamous cell or basal cell carcinoma of the skin in the last two years, self-reported pregnancy, self-reported chronic infection, autoimmune condition or other inflammatory condition, self-reported chronic kidney or liver disease, chronic heart failure or diagnosed with myocardial infarction in the last three months, self-reported diabetes (HbA1c > 8% if known), self-reported acute bacterial or viral infection in the past 24 h or a temperature greater than 38 °C within 24 h of enrolment, self-reported participation in any therapeutic study within 14 days before blood sampling and taking more than 20 mg of prednisone or related drugs.
Clinical blood chemistry was performed on the same samples, including a complete blood count and comprehensive metabolic panel, lipid panel and liver function tests. Basic physical workup (blood pressure, pulse and respirations) was also collected. Lifestyle information was also collected from all participants using a survey which asked about smoking, alcohol, exercise, habits and frequency of consumption of different meats and vegetables.
LonGenity
Details of the LonGenity cohort have been previously published55,56. Briefly, LonGenity is an ongoing longitudinal study initiated in 2008 and designed to identify biological factors that contribute to healthy aging. The LonGenity study enrols older adults of Ashkenazi Jewish descent with an age range of 65–94 years at a baseline. Approximately half of the cohort consists of offspring of parents with exceptional longevity, defined as having at least one parent who survived to 95 years of age. The other half of the cohort includes offspring of parents with usual survival, defined as not having a parental history of exceptional longevity. A total of 962 subjects were included in analyses for this study. The cohort characteristics are summarized in Supplementary Table 1. LonGenity participants are thoroughly characterized demographically and phenotypically at annual visits that include collection of medical history and physical and detailed neurocognitive assessments (described in detail below). The LonGenity study was approved by the institutional review board (IRB) at the Albert Einstein College of Medicine.
Subjects in the LonGenity cohort underwent extensive cognitive examination. The Overall Cognition Composite score was determined by the relative performance of the subject in the Free and Cued Selective Reminding Test, WMS-R Logical Memory I, RBANS Figure Copy, RBANS Figure Recall, WAIS-III Digit Span, WAIS-III Digit Symbol Coding, Phonemic Fluency (FAS), Categorical Fluency, Trail Making Test A and Trail Making Test B. For each task a standardized score (z) was calculated based on the population. The z-score for each task was then combined to create the overall cognition composite.
Stanford Alzheimer’s Disease Research Center
Samples were acquired through the National Institute on Aging (NIA)-funded Stanford Alzheimer’s Disease Research Center (Stanford-ADRC). The Stanford-ADRC cohort is a longitudinal observational study of clinical dementia subjects and age-sex-matched nondemented subjects. The collection of plasma was approved by the Institutional Review Board of Stanford University and written consent was obtained from all subjects. Blood collection and processing were done according to a rigorous standardized protocol to minimize variation associated with blood draw and blood processing. Briefly, about 10 cc of whole blood was collected in a vacutainer ethylenediaminetetraacetic acid (EDTA) tube (Becton Dickinson vacutainer EDTA tube) and spun at 3,000 RPM for 10 mins to separate out plasma, leaving 1 cm of plasma above the buffy coat and taking care not to disturb the buffy coat to circumvent cell contamination. Plasma processing times averaged approximately one hour from the time of the blood draw to the time of freezing and storage. All blood draws were done in the morning to minimize the impact of circadian rhythm on protein concentrations. Plasma pTau-181 levels were measured using the fully automated Lumipulse G 1200 platform (Fujirebio US, Inc, Malvern, PA) by experimenters blind to diagnostic information, as previously described.
All healthy control participants were deemed cognitively unimpaired during a clinical consensus conference that included board-certified neurologists and neuropsychologists. Cognitively impaired subjects underwent Clinical Dementia Rating and standardized neurological and neuropsychological assessments to determine cognitive and diagnostic status, including procedures of the National Alzheimer’s Coordinating Center (https://naccdata.org/). Cognitive status was determined in a clinical consensus conference that included neurologists and neuropsychologists. All participants were free from acute infectious diseases and in good physical condition. A total of 409 subjects were included in analyses for this study. Cohort demographics and clinical diagnostic categories are summarized in Supplementary Table 1.
Stanford Aging Memory Study
SAMS is an ongoing longitudinal study of healthy aging. Blood collection and processing were done by the same team and using the same protocol as in Stanford-ADRC. Neurological and neuropsychological assessments were performed by the same team and using the same protocol as in Stanford-ADRC. All SAMS participants had CDR = 0 and a neuropsychological test score within the normal range; all SAMS participants were deemed cognitively unimpaired during a clinical consensus conference that included neurologists and neuropsychologists. A total of 192 cognitively SAMS participants were included in the present study. The collection of plasma was approved by the Institutional Review Board of Stanford University and written consent was obtained from all subjects. Cohort demographics and clinical diagnostic categories are summarized in Supplementary Table 1.
Knight Alzheimer’s Disease Research Center
The Knight-ADRC cohort is an NIA-funded longitudinal observational study of clinical dementia subjects and age-matched controls. Research participants at the Knight-ADRC undergo longitudinal cognitive, neuropsychologic, imaging and biomarker assessments including Clinical Dementia Rating (CDR). Among individuals with CSF and plasma data, AD cases corresponded to those with a diagnosis of dementia of the Alzheimer’s type (DAT) using criteria equivalent to the National Institute of Neurological and Communication Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association for probable AD58, and AD severity was determined using the Clinical Dementia Rating (CDR)59 at the time of lumbar puncture (for CSF samples) or blood draw (for plasma samples). Controls received the same assessment as the cases but were nondemented (CDR = 0). Blood samples were collected in EDTA tubes (Becton Dickinson vacutainer purple top) at the visit time, immediately centrifuged at 1,500g for 10 min, aliquoted on two-dimensional barcoded Micronic tubes (200 ul per aliquot) and stored at −80 °C. The plasma was stored in monitored −80 °C freezer until it was pulled and sent to Somalogic for data generation. The Institutional Review Board of Washington University School of Medicine in St. Louis approved the study and research was performed in accordance with the approved protocols. A total of 3,075 participants were included in the present study. Cohort demographics and clinical diagnostic categories are summarized in Supplementary Table 1.