Organ-specific aging in the plasma proteome predicts disease

Abstract

In their recent Nature paper, Oh et al. use 4,779 plasma proteins collected across multiple cohorts, publicly available gene expression data, and machine learning models to identify eleven organ-specific aging scores which are linked to organ-specific disease and mortality risk, including heart failure, cognitive decline and Alzheimer’s disease.

Aging is a leading risk factor for many chronic diseases, and previous work indicates that organ systems within the same individual age at differing rates [1]. It therefore follows that quantification of organ-specific aging could be leveraged to predict future health outcomes and reveal novel disease processes, especially for organ-relevant phenotypes. High-throughput, aptamer-based proteomic assays applied to animal and human plasma can be used to train machine learning models for estimating chronological age and calculating so-called ‘age gaps’ (i.e., a measure of an individual’s biological age relative to other same-aged peers based on their molecular profile), which themselves have been associated with mortality risk and other clinically relevant outcomes [2–4]. Oh et al. identified a set of organ-specific proteins in plasma that were used to compute organ-specific age and age gaps, and to test the capacity for such organ-age models to predict risk for age-related, organ-specific diseases. By isolating a set of proteins that were specifically expressed in brain tissue and associated with cognitive function, the authors also identified a brain-specific proteomic signature associated with Alzheimer’s disease (AD) risk and cognitive decline.

Using 4,979 proteins measured across five independent cohorts (n=5,676), the organ-specific proteome was mapped using human organ bulk RNA sequencing data from the Genotype-Tissue Expression project. Cognate genes encoding plasma proteins were classified as mutually exclusive organ-enriched proteins if they were expressed at least four times higher in one organ compared to any other organ, a definition proposed in the Human Protein Atlas (https://www.proteinatlas.org/). After excluding proteins with a high coefficient of variation or a low correlation between different versions of the SomaScan assay across cohorts, 856 (17.9%) proteins were classified as specific to one of eleven different organs: adipose tissue, artery, brain, heart, immune tissue, intestine, kidney, liver, lung, muscle, and pancreas. To determine the optimal combination of proteins for predicting organ-specific biological age, least absolute shrinkage and selection operator (LASSO) regression models were trained on chronological age, with each set of organ-specific proteins as the predictor in separate models. 18.4% of individuals showed extreme age gaps in only one organ (i.e., 2 standard deviations greater difference in organ-specific biological age relative to observed chronological age), while 1.7% showed extreme age gaps in multiple organs, a surprising result given that declining rates of organ function in aging are thought to be highly correlated, especially among older adults.

Higher organ-specific aging gaps were linked to greater disease and mortality risk. For example, heart-aging gaps were associated with greater risk for atrial fibrillation and heart attack, where individuals with atrial fibrillation had hearts that were estimated to be 2.8 years older than their same-aged peers, and individuals with a heart attack had hearts estimated to be 2.6 years older. Interestingly, almost all organ-specific aging gaps were associated with greater risk for certain diseases, such as heart attack and AD. Of the eleven organ-specific aging gaps, ten were associated with future risk of all-cause mortality, with each standard deviation increase in eight age gaps conferring between a 15–50% increased risk. Importantly, multiple organ aging gaps were correlated, albeit only modestly, with a clinical biochemistry-based aging clock (PhenoAge), further suggesting that plasma-based organ-specific aging models capture unique, disease-relevant heterogeneity of aging within and between individuals.

Next, a second generation of cognition-optimized, organ-specific aging models were developed, and their relationship with AD risk was examined. For each set of organ-specific proteins, those that enhanced an organ-age gap’s capacity for predicting cognitive decline were selected using a novel feature permutation algorithm which calculated per-protein associations with both organ-specific biological age gaps and cognitive performance, as measured with the clinical dementia rating (CDR) global score. Using the cognition-optimized subsets of proteins and the same LASSO models implemented for the first generation of organ aging scores, a second generation of cognition-sensitive, organ-specific aging scores was developed. Age gaps for the resulting CognitionBrain, CognitionArtery, CognitionPancreas and CognitionOrganismal (i.e., derived using all proteins which were non-specific to any organ type) were associated with cross-sectional AD risk, and the CognitionArtery and CognitionOrganismal, but not CognitionBrain age gaps, also predicted conversion from cognitively normal to mild cognitive impairment over a 15 year follow up period. In models adjusting for baseline CDR score, plasma pTau-181 (i.e., a leading AD biomarker) and an AD polygenic risk score, the CognitionBrain age gap showed the strongest associations with changes in CDR scores over 5 years, while the combination of the CognitionBrain age gap and plasma pTau-181 was estimated to show an additive effect for predicting CDR changes over this same follow up period. By focusing on proteins used in organ-specific aging models, potential biological mechanisms that may account for the observed associations were postulated. For example, the five proteins used to compute CognitionArtery have been previously implicated in vascular calcification, consistent with prior studies suggesting that vascular dysfunction is an early feature of – and likely a risk factor for – late onset AD [5].

In addition to introducing a framework for modeling organ health and biological aging using plasma proteomics, Oh et al. show their organ aging models can predict mortality, organ-specific functional decline, and organ-specific disease risk, with a particular emphasis on AD. However, several unanswered questions remain, including whether this approach can be applied to other organs not included in the current analyses (e.g., reproductive organs) and to more heterogeneous populations (e.g., non-elderly, non-American, non-Caucasian). Along with these limitations, the authors suggest an important next step should be a direct comparison of their models to existing prediction models (e.g., methylation age clocks). Although several proteins with large weights in organ aging models have been previously implicated in age-related phenotypes (e.g., CPLX 1 and CPLX2 have been implicated in 25-year dementia risk[6,7], and KLOTHO has been implicated in AD [8]) the authors also suggest the need for identifying organ-specific aging proteins that are causal drivers of age-related morbidity. Of course, proteins need not be organ-specific to promote disease, nor do they need to be expressed by the organ through which the disease manifests. Accordingly, these organ age scores may ultimately provide the greatest utility as minimally invasive and scalable biomarkers that can be used as a tool to monitor organ-specific health, particularly before the clinical manifestation of disease or during the course of an intervention. Taken together, the findings by Oh et al. lay the foundation for deconvoluting different rates of clinically relevant aging within individuals and set the stage for measurement of aging at an organ-level resolution.

Figure 1.

Summary of study design. Created with BioRender.com