Organ aging signatures in the plasma proteome track health and disease

Animal studies show aging varies between individuals as well as between organs within an individual1–4, 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.


A. Identification of putative organ-derived plasma proteins
We used the Gene Tissue Expression Atlas (GTEx) human tissue bulk RNA-seq database 16 to identify organ-specific genes and plasma proteins.We determined organ-specificity based on a 4-fold cutoff in bulk RNA-seq data for three main reasons: 1. Determining tissue specificity based on a 4-fold increase in RNA-seq expression ("tissue-enriched") from GTEx and other databases is a well-accepted approach, established by the Human Protein Atlas (HPA) in multiple studies [81][82][83] .The HPA's tissueenriched gene sets are widely trusted and are provided in NCBI, GeneCards, and enrichment analysis tools such as gprofiler 71 .We used the same metric but with the updated, more deeply sequenced GTEx RNA-seq dataset and with a more generalizable framework for tissue->organ mapping (Supplementary Table 2).
2. We considered determining organ-specificity based on tissue protein levels from a human tissue proteomics atlas (Jiang et al) 84 ; however, we opted not to because organ protein levels may be misleading in regard to determining the original organ source of the protein.Specifically, a protein may be present in an organ because it was trafficked there after being synthesized by another organ and secreted into the plasma.Albumin and complement proteins are not enriched at the protein level in the liver even though they are synthesized there, and there are proteins which are synthesized in the hypothalamus that are enriched in the pituitary because they are stored there before release 84 .Generally, discordance between protein and RNA levels are interpreted as a result of protein trafficking/export/secretion, while enrichment at the RNA level is recognized as the tissue of origin for protein synthesis 81,82,84,85 .It may also be true that proteins which are present at the protein level in an organ but are not synthesized there also contain important information about said organ.We believe this idea of cross-organ communication in aging is an exciting area for future study.For the current manuscript, our goal was to determine the putative organ source of plasma proteins to infer organ age.
3. RNA-seq data contains nearly full coverage of the genome, while proteomics data has much lower coverage.In Jiang et al, only 6320 proteins were detected in >50% of samples, and these are heavily biased towards abundant proteins, which are detectable by mass spectrometry.The percentage of these mappable to the SomaScan plasma proteomics assay is even lower.Determining organ-specificity based on RNA-seq data increased our coverage of the mappable organ-specific plasma proteome.
We then compared every non-zero group with the zero group (denoting the non-zero group as 1 and the zero group as 0) for changes in mortality risk.We did this analysis for each of the aging models.We did not adjust for multiple comparisons because the assumptions were not met: each statistical test is done in a different subset of individuals, and tests for different bins in the same organ are generally correlated.
Interestingly, the association between the age gap and mortality risk was non-linear for some organs, such as the heart, brain, pancreas, kidney.The relationship with the heart age gap seems to be U-shaped where both high (+1, +2, +3) and extremely low heart age gaps (-2) are associated with increased mortality risk.The kidney age gap was also interesting in that it was not associated with mortality risk when looking at the whole age gap distribution (Fig. 2j), but the +3 age gap group was positively associated with mortality, suggesting the "extreme agers" framework may be more useful for certain organs and traits.Other organs, including the organismal, adipose, artery, and immune, show a more linear relationship with mortality risk.
Whether these nonlinear dynamics also exist for other aging biomarkers, such as methylation clocks, is unknown.This analysis points to a need for additional studies on the relationship between extreme aging and disease risk.

C. Relationships between blood biochemistry markers and organ aging.
While a full analysis of all clinical biochemistry markers is challenging, there are a number of additional interesting relationships in the data.
• BUN: Kidney, adipose, brain, immune, and muscle age gaps are significantly positively associated with BUN, artery age gap is significantly negatively associated.The strongest association is with the kidney age gap.While BUN is not specific, it is often considered a marker of kidney function clinically.
• AST: Kidney, heart, and artery age gaps are positively significantly associated with AST, while brain is significantly negatively associated.AST variation within the normal range is difficult to interpret clinically.Abnormally high AST is often a sign of liver or heart disease, and moderately high AST is most often noted as a sign of elevated cardiovascular risk in middle aged and elderly populations.
• ALT: The brain, control, liver, intestine, kidney, organismal, and pancreas age gaps are significantly negatively associated with ALT, while the kidney age gap is significantly positively associated with ALT.PhenoAge gap is positive but not significant.As discussed in the text, this is yet another U-shaped aging biomarker.Low ALT in the elderly is associated with increased frailty and reduced survival and has been previously suggested as a biomarker of aging 86 .Abnormally high ALT can be a marker of acute liver damage, although it is also produced by other tissues and is not specific.
• Albumin: The immune, heart, liver, organismal, control, and PhenoAge gaps are significantly negatively associated with albumin levels.The strongest association is with the liver age gap.Albumin is produced by the liver, although it is not detected by the SomaScan assay so it is not a protein in the liver aging model.Clinically, lower albumin could be considered as a sign of worse health, and it can be low in a number of liver, kidney, and digestive diseases as well as in malnutrition/undernutrition.
• Plasma glucose is significantly positively associated with PhenoAge age gap and kidney age gap, while intestine and liver age gap are significantly negatively associated.The strongest association is with PhenoAge, which is unsurprising since plasma glucose is the highest weighted input biomarker in the PhenoAge model.Both kidney and intestine age gap are positively associated with diabetes incidence but have differential associations with plasma glucose.This further supports the hypothesis that different organ models could be measuring different aspects of aging, in this case metabolic aging.Insulin resistance, glucose response, and glucose levels are all known to degrade with age, but insulin levels and glucose response have been noted to change more dramatically than fasting blood glucose level 87 .
There are many biomarkers of health which have a nonlinear relationship to aging outcomes, and in the elderly many relationships between biomarkers and health/mortality/frailty reverse direction compared to young and middle-aged adults.The distribution and mean age of the population that an aging model is trained on will thus impact associations with traits.This is not frequently discussed or accounted for in models of molecular aging.Such a case is illustrated by diastolic blood pressure, where the strongest association was with heart aging (adjusted Pearson r=-0.18,q=2.62e-10).Nine organ age gaps (adipose, brain, control, heart, intestine, kidney, liver, muscle, organismal, pancreas) were significantly associated with decreases in diastolic blood pressure, while the opposite association was seen with the PhenoAge age gap (Supplementary Fig. 5a, Supplementary Table 14).Diastolic blood pressure was one of many traits with a U-shaped relationship to aging outcomes (Supplementary Fig. 5b).While high blood pressure in young and middle-aged adults is indicative of cardiometabolic dysfunction, in the elderly low blood pressure is common and more strongly associated with mortality and frailty [88][89][90] , though high blood pressure is also detrimental 91 .The differences between PhenoAge and the organ age models could be due to differences in the age distribution of the underlying training cohorts for the models.Our models were trained in the KADRC, which has a greater proportion of elderly individuals, while PhenoAge was trained in NHANES, which has a greater proportion of young individuals.This kind of U-shaped relationship with age and aging outcomes is quite common and is also seen with BMI 92 .Prospective studies in older adults have shown that while obesity slightly increases mortality and cardiovascular disease risk, the highest risk groups are those with a BMI under 23.Interestingly, the intestine and pancreas age gaps show a negative association with BMI and obesity but a positive association with mortality risk, while the kidney age gap shows a positive association with BMI, suggesting that the full picture of organ health in aging and disease may be more complex than currently understood.

D. Relationship between CognitionBrain age gap and brain volume.
To further examine the relationship between the CognitionBrain age model and brain aging, we tested associations between CognitionBrain age gap and changes in brain volume.We used plasma-matched brain MRI data from 469 individuals in the Stanford-ADRC and SAMS cohorts to assess the relationship between the CognitionBrain age gap and brain region-specific volumes (Extended Data Fig. 7c, Supplementary Table 22).39 out of 65 (60%) associations were significant after multiple hypothesis correction.The most significant associations were negative associations with the superior frontal cortex (adjusted r=-0.20,q=8.49e-5), hippocampus (adjusted r=-0.21,q=1.36e-4), and total cortex (adjusted r=-0.20,q=1.39e-4), whereby individuals with smaller brain region volumes appeared older based on their CognitionBrain age gaps.We also found a negative association with the AD signature region (adjusted r=-0.16,q=3.61e-3), a composite measure of the parahippocampal gyrus, entorhinal cortex, inferior parietal lobes, hippocampus, and precuneus 93 .
We then compared our plasma proteomics-based brain age to two MRI brain aging clocks.
Based on its established publication record, we started with the BARACUS model 78 , a linear support vector machine based aging clock trained on brain MRI-based volumetric data from 1,166 cognitively normal individuals aged 20-80.However, when assessing predicted versus chronological age correlation, we noticed an odd technical artifact: the predicted age had a ceiling near 75, even for individuals with chronological age above 90.Looking more closely at the original publication, we found the same issue of an upper ceiling, and also a lower ceiling, to predicted age.This leads us to believe that the BARACUS algorithm cannot accommodate all ages in our cohort.
Due to this technical limitation of BARACUS, we also assessed brainageR 14 , a Gaussian Processes based aging clock trained on brain MRI-based volumetric data from n=3,377 cognitively healthy individuals aged 18-92, and which has shown better performance than BARACUS in other studies 94 .The CognitionBrain age gap was positively correlated with the brainageR age gap (r=0.16,p=7.51e-4) (Extended Data Fig. 6h), but not as strongly as the correlation between CognitionBrain age gap and individual brain volumes (ie.hippocampus: adjusted r=-0.21,q=1.36e-4).This is likely due to the fact that BARACUS and brainageR do not take into account total intracranial volume and thus capture more noise.
E. Literature review of highly weighted brain aging proteins.