Long-term temporal stability of circulating proteins in older adults

Abstract

Circulating proteins provide insights into an individual’s health and aging trajectory. However, little is known about their long-term temporal stability. We evaluated the five-year temporal stability of 7288 serum proteins in 3093 participants (mean age 75 years) of the AGES-Reykjavik study. We observe wide variability in protein temporal stability, independent of transcriptomic stability in tissues. Temporally stable proteins tend to be extracellular, secreted, and tissue-specific, while temporally variable proteins are typically involved in intracellular housekeeping functions. Generally, protein levels become more stable with age, and for some proteins, temporal stability differs by sex. Similar patterns and consistent temporal stability were observed for a subset of plasma proteins in the Cardiovascular Health Study. Genetic factors and disease state have widespread effects on the temporal stability of the proteome, while the co-regulatory structure of the proteome is largely retained over time. Our findings underscore the protein-specific differences in long-term temporal stability, which are particularly important to consider in the context of biomarker development and precision medicine.

The circulating proteome offers insights into an individual’s health and aging. Here, the authors show that the long-term temporal stability of circulating proteins is linked to their distinct biological properties and is influenced by age, genetics and disease state, providing an important reference for longitudinal biomarker studies.

Introduction

The circulating proteome refers to all proteins present in serum or plasma, including actively secreted proteins as well as proteins entering the circulation through membrane shedding, extracellular vesicles (exosomes), cell turnover or damage, or other physiological release mechanisms¹. It has been intensively studied to understand how protein levels vary between individuals in relation to health and disease, and as a source of novel biomarkers². Recent technological advances have enabled high-throughput measurement of thousands of circulating proteins simultaneously, uncovering links between protein abundance and genetic variation, disease risk, and disease progression². For example, we and others have highlighted the strong influence of genetics on circulating protein levels^3–7, and proteomic profiling has revealed widespread associations between protein abundance and common diseases^8–15. We have previously shown that the circulating proteome is organized into co-regulated modules linked to multiple disease states⁷, and key contributing factors to plasma protein variability have been shown to include kidney and liver function, smoking, and inflammation⁶. While collective emphasis has been on using large cohorts to study various disease states, less focus has been put on how protein levels fluctuate within and between individuals over time. Understanding both inter- and intra-individual variability is particularly important in the context of precision medicine, which aims to tailor prevention and treatment strategies to each person’s unique characteristics¹⁶.

Protein levels have been shown to change with aging¹⁷, motivating the development of proteomic aging clocks to estimate chronological age and predict mortality risk^18,19. Longitudinal studies have also revealed nonlinear patterns in proteomic aging, including distinct waves of change in midlife and later life^17,20. Some longitudinal studies have also shown that the circulating proteome as a whole is mainly stable within individuals but displays protein-specific variability. For instance, three studies using different plasma proteomic technologies (Olink²¹, mass spectrometry²² (MS), or multiplexed antibody assays²³) consistently found that proteomic profiles are highly individual-specific and generally stable, with greater variability between individuals than within the same individuals over time (1-2 years). Stability, however, varied across proteins²¹ and depended on definitions and measurement approaches. Studies integrating proteomic and genomic data further revealed strong genetic contributions to these stable patterns²⁴. Collectively, these studies demonstrate the importance of identifying factors that influence the dynamics of the circulating proteome that can mirror the real-time global health state of individuals. Despite these insights, prior work has been constrained by small cohorts, short follow-up periods, and limited protein coverage. Furthermore, it is unknown how the temporal stability of proteins relates to disease onset in aging.

Here, leveraging over 7000 serum protein measurements for 3093 participants from two visits of the AGES-Reykjavik study (Fig. 1), we show that long-term temporal stability of serum protein levels varies widely and is not determined by the magnitude of change in protein abundance. We validate a subset of these findings for plasma proteins measured with a different analytical platform in the Cardiovascular Health Study (CHS). We find that proteins with differing temporal stability exhibit distinct biological characteristics and that genetic effects and disease states influence protein behavior over time. Together, these findings highlight the importance of accounting for temporal variation in protein levels, particularly when considering proteins as biomarkers.

Fig. 1. Overview of the study.

The study included participants from the AGES-Reykjavik cohort with serum protein measurements at both baseline and follow-up visits ( ~ 5 years apart) and available genotype data (n = 3093; mean age = 75 years). Approximately 7000 human serum proteins were measured at both time points using the SomaScan 7k (v4.1) platform. The validation cohort consisted of a subset of the Cardiovascular Health Study (CHS), in which plasma proteins were measured at two visits ( ~ 4 years apart) using the Olink Explore HT platform (n = 2770; mean age = 72 years). Temporal stability of each protein was defined separately in AGES and CHS using the Spearman correlation coefficient (SCC) between baseline and follow-up protein levels. Proteins were classified as temporally stable if individuals showed similar levels across visits (SCC > 0.75), and as temporally variable if levels differed between visits (SCC < 0.4). Temporal stability was examined in relation to functional characteristics of proteins, demographic factors, genetic effects, and disease status relative to study visits across multiple conditions. Created in BioRender. Gudmundsdottir, V. (https://BioRender.com/2yrxow2).

Results

Temporal stability of circulating proteins

The current study included 3093 AGES participants aged 66–93 years (mean 75 years), of whom 1794 (58%) were women. Each individual had serum measurements for 7288 human proteins (SomaScan v4.1) available from two study visits on average 5.2 years apart (Supplementary Data 1 & 2), and genotype data available⁴.

Temporal stability of serum proteins between time points was assessed using the Spearman correlation coefficient (hereafter referred to as temporal SCC). Temporal stability varied across proteins, as indicated by a wide distribution of temporal SCC estimates, with a median value of 0.58 (Fig. 2A). Based on suggested guidelines for the interpretation of SCC values²⁵, we defined the temporally stable proteins (n_prot = 729) as having SCC greater than 0.75 and temporally variable proteins (n_prot = 681) less than 0.40, which closely approximated the 90^th and 10^th percentiles of the distribution (0.75 and 0.41, respectively) (Fig. 2A; Supplementary Data 3). Proteins with intermediate SCC values (0.4–0.75) were classified as having medium temporal stability. This approach ensures that the stable and variable groups represent proteins with clearly distinct temporal behavior, while avoiding arbitrary inclusion of borderline cases. The most temporally stable protein was LILRA6 (temporal SCC = 0.97; Fig. 2A), a leukocyte immunoglobulin-like receptor A6 that functions as a receptor for MHC I antigens²⁶. An example of a protein with medium temporal stability was SERPINA3 (serpin family A member 3; temporal SCC = 0.56; Fig. 2A), a serine protease inhibitor²⁷. Finally, an example of a temporally variable protein was CD2AP (CD2-associated protein; temporal SCC = 0.29; Fig. 2A), a regulator of signal transduction and cytoskeletal organization²⁸.

Fig. 2. Temporal stability of circulating proteins.

The temporal stability of proteins was assessed using temporal SCC estimates in the AGES cohort (n = 3093). Proteins were divided into three temporal stability groups: variable ( < 0.40, n_prot = 681), medium (0.40–0.75, n_prot = 5855), and stable ( > 0.75, n_prot = 752). The protein temporal stability was validated using external Olink plasma protein measurements from 2770 participants in the Cardiovascular Health Study (CHS). A Distribution of temporal SCC estimates of all SOMAmers (n_prot = 7288, median = 0.58) measured in 3093 AGES serum samples. Examples of proteins from each temporal stability group are visualized with scatterplots of baseline (AGES) versus follow-up (AGESII) protein levels for each individual. LILRA6 was the most temporally stable protein (temporal SCC = 0.97), SERPINA3 showed medium temporal stability (temporal SCC = 0.56), and CD2AP was temporally variable (temporal SCC = 0.29). B Distribution of temporal SCC estimates using Olink plasma protein measurements from CHS after filtering out proteins with low platform correlation (SCC < 0.4) between SomaScan and Olink (n_prot = 1301, median = 0.63). C Distribution of temporal SCC estimates in CHS using Olink stratified by temporal stability groups (stable, n = 304; medium, n = 903; variable, n = 94) defined in AGES (n_prot = 1301). Four asterisks (****) indicate a significant difference in distribution between groups (Wilcoxon rank-sum test, two-sided, P < 2 × 10⁻¹⁶). Boxplots indicate median value, 25^th, and 75^th percentiles. Whiskers extend to the smallest/largest value of no further than 1.5 $\times$ interquartile range. Outliers are not shown.

To assess whether the temporal stability patterns observed in AGES were reproducible, we performed validation analyses in the Cardiovascular Health Study (CHS)²⁹ (mean age 72 years, 58% women) using 5420 plasma protein measurements from the Olink Explore HT in 2770 participants at two study visits, on average 4 years apart (Supplementary Data 4). Validation was restricted to 1301 proteins with at least moderate platform correlation (SCC > 0.4) between SomaScan and Olink in CHS at baseline. As in AGES, temporal SCC estimates in CHS showed a wide distribution (median = 0.63) (Fig. 2B). While the distribution in CHS appeared more bimodal compared with the approximately normal distribution in AGES, we observed good correlation (SCC = 0.68, P = 6.36 × 10⁻¹⁷⁶) between temporal stability estimates from the two cohorts (Supplementary Fig. 1). Temporal SCC in CHS differed across temporal stability categories defined in AGES (Kruskal–Wallis test, χ² = 389.78, df = 2, P = 2.29 × 10⁻⁸⁵), with proteins classified as temporally variable in AGES (n_prot = 94) having a lower median SCC in CHS (0.43) than those classified as temporally stable (0.79, n_prot = 304) (pairwise Wilcoxon rank-sum test, two-sided, FDR = 4.35 × 10⁻⁴²), suggesting good agreement in temporal stability between the two cohorts (Fig. 2C).

Technical factors influencing protein temporal stability

Absolute and relative change in protein levels over time is of interest in longitudinal biomarker studies. However, the temporal SCC showed almost no correlation with the absolute difference in protein measurements (median delta value) between time points (SCC = 0.05, P = 3.90 × 10⁻⁶; Supplementary Fig. 2). Although statistically significant, the effect size was negligible, suggesting that temporal SCC captures a complementary aspect of protein temporal stability rather than reflecting absolute changes in protein levels over time. However, there was a moderate relationship between the intra-individual variability (temporal SCC) and inter-individual variability of proteins (defined as inter-quartile range), both at baseline (AGES) and the follow-up (AGESII) visit (SCC AGES = 0.22, P = 3.76 × 10⁻⁸³; SCC AGESII = 0.26, P = 6.86 × 10⁻¹¹⁷), where proteins with higher intra-individual variability were also more likely to have high inter-individual variability (Supplementary Fig. 3).

The SomaScan protein measurements are provided in relative fluorescent units (RFU). Median RFU differed across protein temporal stability groups (Kruskal–Wallis test, χ² = 318.59, df = 2, P = 6.59 × 10⁻⁷⁰). Pairwise comparisons revealed that the stable proteins had significantly higher RFU values than those in the medium (pairwise Wilcoxon rank-sum test, two-sided, FDR = 5.64 × 10⁻⁶⁹), or variable groups (pairwise Wilcoxon rank-sum test, two-sided, FDR = 4.02 × 10⁻¹⁹; Supplementary Fig. 4). While RFU values are not directly comparable between proteins nor can they be interpreted as a measure of abundance, we did find a strong correlation between RFU values and external blood concentration estimates from the Human Protein Atlas (SCC = 0.71, P = 1.52 × 10⁻⁸⁹, n_prot = 569; Supplementary Fig. 5). Thus, the higher RFU values observed for temporally stable proteins could indicate that they are generally of higher abundance, although we cannot exclude the possibility that other technical factors contribute to the observed RFU differences.

To estimate whether our temporal stability results reflected technical variability, we extracted inter-plate percentile variation (PV) values from a recent reproducibility study of the SomaScan v4.1 platform conducted in the Atherosclerosis Risk in Communities (ARIC) Study³⁰. We compared these PV values with the temporal SCC value in AGES for each protein. As could be expected, proteins with lower temporal SCC estimates in AGES tended to have higher PV in ARIC (SCC = −0.33, P = 1.47 × 10⁻¹⁸⁶; Supplementary Fig. 6), indicating more technical variability among the proteins defined as temporally variable. Roughly 82% of the temporally variable proteins had a PV < 14.11% (corresponding to Z-scores $\le$ 2) in ARIC, indicating that they are technically robust. Proteins with PV > 14.11% (Z-score > 2) were excluded from downstream analyses to reduce the effect of technical variability when comparing stable and variable proteins, leaving 7097 proteins for analysis (median SCC = 0.61) (Supplementary Data 5).

Functional characteristics of temporally stable and variable proteins

To evaluate biological characteristics of the temporal stability protein groups, we performed functional enrichment analysis using all 7097 technically stable SomaScan 7k proteins as background. Unless otherwise specified, all enrichment results reported passed multiple testing correction (FDR < 0.05). Overall, a clear functional separation was observed between the stable and variable protein groups when incorporating gene ontology (GO) terms, KEGG, WikiPathways, and Reactome (Fig. 3A; Supplementary Data 6), indicating a biological distinction relating to temporal stability. Only a few terms were significantly associated with both stable and variable proteins, including the immune system and the secretory granule. The terms “extracellular region” and “extracellular space” were also significantly enriched in both temporal stability groups, though the enrichment was much stronger among stable proteins.

Fig. 3. Enrichment and functional analysis of temporally stable and variable proteins.

A Comparison of enriched pathways from g:Profiler between temporally stable and variable proteins, demonstrating distinct biological roles with only a few pathways shared across groups. Data points are colored based on whether temporally stable, variable proteins, or both are significantly enriched within the pathway. B Top 15 enriched Gene Ontology (GO) driver terms and (C) enriched pathways from KEGG, Reactome, and WikiPathways for temporally stable proteins. D Top 15 GO driver terms and (E) enriched pathways from KEGG, Reactome, and WikiPathways for temporally variable proteins.

Among significantly (two-sided Fisher’s exact test, FDR < 0.05) enriched GO driver terms for the temporally stable proteins, the strongest enrichment included extracellular region (FDR = 7.18 × 10⁻¹⁸), cell periphery (FDR = 2.40 × 10⁻¹⁰), and signaling receptor activity (FDR = 5.99 × 10⁻¹⁰; Fig. 3B, Supplementary Data 7). The strongest enriched pathway was “Immunoregulatory interactions between a lymphoid and a non-lymphoid cell” (FDR = 9.92 × 10⁻¹⁰), followed by immune system (FDR = 3.78 × 10⁻³), viral protein interaction with cytokine and cytokine receptor (FDR = 4.85 ×  10⁻³), salivary secretion (FDR = 4.85 × 10⁻³) and the complement system (FDR = 2.31 × 10⁻²; Fig. 3C, Supplementary Data 7). Thus, temporally stable proteins tend to have established extracellular roles, particularly in signaling and immune pathways.

In contrast, the temporally variable proteins were most strongly enriched (two-sided Fisher’s exact test, FDR < 0.05) for the GO driver terms cytosol (FDR = 1.13 × 10⁻¹¹), purine ribonucleoside triphosphate binding (FDR = 3.28 × 10⁻¹⁰), and GTPase activity (FDR = 3.28 × 10⁻¹⁰; Fig. 3D, Supplementary Data 8). The top enriched pathways included RAB geranylgeranylation (FDR = 3.55 × 10⁻⁸), platelet activation, signaling and aggregation (FDR = 8.10 × 10⁻⁸), membrane trafficking (FDR = 2.34 × 10⁻⁶), vesicle-mediated transport (FDR = 4.85 × 10⁻⁶) and cell cycle (FDR = 2.71 × 10⁻⁵; Fig. 3E, Supplementary Data 8). Overall, the temporally variable proteins are thus more likely to participate in intracellular functions related to cell cycle regulation and intracellular trafficking. Finally, we performed the same enrichment analysis for temporal stability protein groups independently defined in the CHS cohort, which revealed similar biological signals (Supplementary Data 9 & 10).

Mirroring the patterns observed from GO term and pathway memberships, the temporal stability protein groups were enriched for different protein classes, as defined by the Human Protein Atlas (Fig. 4A; Supplementary Data 11). The stable proteins (n_prot = 649) were significantly (two-sided Fisher’s exact test, FDR < 0.05) enriched for predicted secreted proteins (FDR = 5.32 × 10⁻³²), predicted membrane proteins (FDR = 2.17 × 10⁻¹¹), and CD markers (FDR = 1.23 × 10⁻¹³), while the variable proteins (n_prot = 541) were enriched for candidate cardiovascular disease genes (FDR = 3.21 × 10⁻²), predicted intracellular proteins (FDR = 9.28 × 10⁻⁴), cancer-related genes (FDR = 7.19 × 10⁻³), and enzymes (FDR = 2.92 × 10⁻³). Both groups were, however, enriched for plasma proteins. Stable proteins were more likely to be encoded by tissue-enriched genes (FDR = 1.88 × 10⁻¹⁸), and to exhibit tissue-enriched protein expression (FDR = 9.78 × 10⁻¹⁵), whereas temporally variable proteins were under-represented in both categories (FDR = 3.22 × 10⁻⁷; FDR = 5.92 × 10⁻¹⁶; Supplementary Data 12). Similar patterns were observed when considering particular organ- and tissue-specific expression profiles. Gene expression profiles defined from GTEx transcriptomic data³¹ revealed that stable proteins (n_prot = 722) were consistently more likely to exhibit organ-restricted expression, particularly in adipose (FDR = 1.91 × 10⁻²), immune (FDR = 2.73 × 10⁻¹¹), liver (FDR = 2.53 × 10⁻²), pancreatic (FDR = 6.55 × 10⁻³), and prostate tissues (FDR = 3.59 × 10⁻²), while variable proteins (n_prot = 543) were only enriched among genes with no tissue-enriched expression (FDR = 6.15 × 10⁻³; Fig. 4B). Similar patterns were observed at different definitions of organ-specificity and at the tissue-level (Supplementary Data 13 & 14), with stable proteins consistently more likely to exhibit organ-restricted expression, particularly in immune-related tissues, compared to variable proteins (Supplementary Fig. 7).

Fig. 4. Functional characteristics of temporally stable (n = 729) and variable (n = 681) proteins.

A Protein classes and tissue-enriched gene/protein expression across temporal SCC groups (data available for 649 stable and 541 variable proteins). Forest plots show odds ratios (two-sided Fisher’s exact test; log-scaled x-axis) with 95% confidence intervals (error bars); filled symbols indicate FDR < 0.05. B Organ-specificity (defined as ≥2-fold higher expression in the highest-expressed organ compared to the second highest) across temporal SCC groups (data available for 722 stable and 543 variable proteins), based on external gene expression data. Odds ratios are presented as in (A). C Proportion of proteins with loss-of-function (LoF) intolerance (pLI ≥ 0.9) across temporal stability groups (data available for 647 stable and 540 variable proteins). Enrichment was assessed using two-sided Fisher’s exact test (***, P < 1 × 10⁻³). D Degree (log-scaled) in the protein–protein interaction (PPI) network across temporal stability groups (data available for 291 stable and 277 variable proteins), and degree (log-scaled) in the co-regulatory serum network across temporal stability groups (data available for 457 stable and 331 variable proteins). For boxplots, center lines indicate the median; boxes show the 25th and 75th percentiles; whiskers extend to 1.5 × the interquartile range; outliers are not shown. Differences between groups were assessed using two-sided Wilcoxon rank-sum tests (****, P < 1 × 10⁻⁴).

Finally, temporally variable proteins had a higher proportion of proteins encoded by genes intolerant to loss-of-function (LoF) variants (24% of 540 variable proteins) compared to the stable ones (11% of 647 stable proteins) (Fig. 4C). Additionally, the variable proteins had significantly higher hub status in both a physical protein-protein interaction (PPI) network³² (Fig. 4D; Wilcoxon test, two-sided, P = 3.27 × 10⁻¹⁷, n_stable = 291, n_variable = 277) and the serum co-regulatory network⁷ (Fig. 4D; Wilcoxon test, two-sided, P = 4.18 × 10⁻²⁷, n_stable = 457, n_variable = 331). Thus, temporally variable proteins tend to be network hubs under stronger evolutionary constraint than those that are temporally stable in the circulation.

Gene expression variance and protein temporal stability

The enrichment results above demonstrate that temporally variable proteins are generally more involved in housekeeping processes than stable proteins. This finding was somewhat unexpected, as our initial hypothesis was that housekeeping proteins would have stable expression and, thus, be temporally stable in the longitudinal serum proteomics data. We therefore explored gene expression variability over time and its relationship with circulating protein temporal stability to evaluate if this pattern would also be reflected at the transcriptomic level and across different tissues.

We utilized data from a study examining the landscape of gene expression variance across 57 RNA-sequencing datasets from various tissues and distinct populations³³. In this study, gene expression variance was calculated as the standard deviation (SD) of gene expression levels for each gene across all samples, with each gene ranked based on this value. We observed a significant difference in gene expression variance between the temporally stable (n_prot = 339) and variable protein groups (n_prot = 482, Wilcoxon rank-sum test, two-sided, P = 1.74 × 10⁻²¹; Supplementary Fig. 8). Specifically, the stable proteins had higher SD rank values, indicating more gene expression variance overall, whereas variable proteins had lower gene expression variance. This suggests an inverse relationship between gene expression variance and protein temporal stability. Overall, this observation indicates that temporal stability in serum may only partially capture the transcriptomic variance across tissues and individuals and may instead reflect the mechanisms by which the proteins enter, or are cleared from, the circulation.

Protein temporal stability across demographics

To evaluate whether protein temporal stability varied across age and sex, we performed a linear regression analysis with an interaction term. Among 7097 proteins tested, 1164 showed significant age-dependent effect modification (FDR < 0.05). These interactions were more likely to be observed among temporally variable proteins (n_variable = 160, n_stable = 68) and most were positive (1094 positive vs. 70 negative), indicating that temporal stability tends to increase with age (Fig. 5A; Supplementary Data 15). In separate cross-sectional analyses, protein associations with age were assessed independently at baseline and at follow-up (Supplementary Data 16 & 17). In these models, 186 proteins were not significantly associated with age at either time point but showed significant (FDR < 0.05) age-dependent modification of temporal stability. This suggests that temporal protein stability may capture age-dependent dynamics that are not apparent from cross-sectional measurements at a single time point.

Fig. 5. Temporal stability across demographics.

Modification of temporal stability by age (A–C) or sex (D–F) in AGES (n = 3093). Volcano plots (A, D) indicate interaction terms between protein and age (A) or sex (D) on the x-axis and -log10(FDR) for the interaction term on the y-axis. Interaction terms were tested using linear regression models. Proteins with significant interactions after multiple testing correction (FDR < 0.05) are colored by direction of effect: red indicates a positive interaction (greater temporal stability with age or in females), and blue indicates a negative interaction (reduced temporal stability). Non-significant proteins are shown in gray. Proteins available for replication in CHS are outlined and colored in dark red/blue if directionally consistent and nominally significant (P < 0.05), in light red/blue if directionally consistent but P > 0.05 and otherwise shown in gray. Proteins with significant (FDR < 0.05) interactions in the same direction in both cohorts are labeled in bold black. A Age-modified temporal stability effects for all proteins (n = 7097). Replicated proteins include SIGLEC5 and EFNA1. B–C Relationship between baseline and follow-up protein levels for SIGLEC5 (B) and SOD1 (C) in AGES (n = 3093), stratified by age ( ≥ 75 years at baseline, red; <75 years, blue). Regression slopes for each group are shown. Boxplots of protein levels at each visit are displayed. D Sex-modified temporal stability effects for all proteins (n = 7097). Replicated proteins include SPINT3, PSPN, PNP (two SOMAmers), and DSCAM. E–F) Relationship between baseline and follow-up protein levels for SPINT3 (E) and SPP1 (F) in AGES (n = 3093), stratified by sex (females red, males blue). Regression slopes for each group are shown. Boxplots of protein levels at each visit are displayed. B–C, E–F Boxplots show the median and interquartile range; whiskers extend to 1.5× IQR. Group differences were tested using the two-sided Wilcoxon test and indicated by asterisks (ns, not significant; **, P < 1 × 10⁻²; ****, P < 1 × 10⁻⁴). Outliers are not shown.

Of the 1164 serum proteins with significant (FDR < 0.05) age-dependent modification of temporal stability in AGES, 193 could be evaluated for replication in the CHS cohort (Supplementary Data 15). The age-interactions for two proteins, SIGLEC-5 (sialic acid–binding Ig-like lectin 5) and EFNA1 (ephrin-A1), were replicated (FDR < 0.05) and additional 15 proteins showed suggestive (P < 0.05) validation with consistent direction of effect (Fig. 5A). EFNA1 is a cell-surface ephrin ligand that mediates cell–cell communication through Eph receptor signaling and has been implicated in tumor progression³⁴ (Supplementary Fig. 9). SIGLEC-5 is a member of the CD33-related Siglec family that regulates immune cell signaling³⁵. Its levels were increased with age at both AGES visits (Wilcoxon rank-sum test, two-sided, P_baseline = 2.81 × 10⁻⁶, P_follow-up = 1.52 × 10⁻⁹), which could partly drive the increase in stability with age (Fig. 5B). Another example is SOD1 (superoxide dismutase 1), which was among the top age-interaction proteins in AGES and showed suggestive replication in CHS (nominal P < 0.05; Fig. 5C). This protein is an antioxidant, but its post-translational modifications have also been implicated in the aging process³⁶. Similarly to SIGLEC-5, the regression slope between baseline and follow-up SOD1 levels was lower in participants younger than 75 years old compared to those older (0.54 vs 0.69, interaction FDR = 1.35 × 10⁻⁶), suggesting that SOD1 temporal stability also increases with age (Fig. 5C).

A similar analysis was performed to assess sex-dependent effects on temporal stability, yielding 1081 proteins with support for sex-dependent effect modification (FDR < 0.05; Fig. 5D; Supplementary Data 18). Of these, 1007 (93%) proteins showed lower temporal stability in women than in men. Similar to age, 283 proteins were not associated with sex at either time point in cross-sectional models but had a significant (FDR < 0.05) interaction with sex (Supplementary Data 16 & 17). In CHS, 163 proteins could be evaluated for replication and the interaction with sex was replicated for five proteins (FDR < 0.05; Fig. 5D), while nine additional proteins showed suggestive validation with consistent effects (Fig. 5D; Supplementary Data 18). SPINT3 (Kunitz-type protease inhibitor 3) was one of the replicated proteins and showed significantly higher levels in males than females at both baseline and follow-up visits (Wilcoxon rank-sum test, two-sided, P < 2.2 × 10⁻¹⁶; Fig. 5E). Comparing the regression slope between baseline and follow-up protein levels, the slope was closer to unity in males than in females (0.91 vs 0.68, interaction FDR = 9.33 × 10⁻¹⁷), indicating greater temporal stability in males (Fig. 5E). This likely reflects the fact that SPINT3 is highly expressed in male reproductive tissues, both at the RNA and protein levels³³, and therefore the lower temporal stability observed in females may reflect the fact that SPINT3 is not normally expressed at high levels, resulting in less temporal stability over time.

Another protein with suggestive validation in CHS (P < 0.05) was SPP1 (osteopontin). Although SPP1 levels did not differ significantly between sexes at either visit (Wilcoxon rank-sum test, two-sided, ns), temporal stability was lower in females, with a smaller regression slope between baseline and follow-up levels compared with males (0.58 vs 0.66, interaction FDR = 8.87 × 10⁻³) (Fig. 5F). Given that SPP1 is a secreted inflammatory and tissue-remodeling protein whose expression can be influenced by hormonal and immune signaling³⁷, this lower temporal stability in females may reflect sex-specific regulatory effects over time.

Genetic effects on temporal stability

We have previously described the genetic effects on proteomic measurements in the AGES study^4,38. In total, 6,572 SOMAmers had at least one conditionally independent and genome-wide significant (P < 5 × 10⁻⁸) protein quantitative trait locus (pQTL). Among these, 278 SOMAmers had only cis-pQTLs (residing within 500 kb of the respective gene), 1,468 had both cis- and trans-pQTLs, and 4826 had only trans-pQTLs. The remaining 716 SOMAmers had no significant pQTLs.

To assess whether genetic signals influence protein levels over time, protein measurements previously considered as technically robust (n_prot = 7097; see “Methods”) were adjusted for conditionally independent pQTLs, thereby removing genotype variability in the protein measurements. The temporal SCCs were then recalculated (hereafter referred to as adjusted temporal SCC) to assess any shifts in temporal stability upon the removal of genetic drivers (Supplementary Data 19). The comparison between original and adjusted temporal SCC values (Fig. 6A) showed a strong overall correlation (SCC = 0.93), and the median SCC changed from 0.61 to 0.59 (paired t-test, two-sided, P = 2.14 × 10⁻²³³). SOMAmers with cis-pQTLs had higher temporal SCC than other proteins, both before and after adjustment (Supplementary Fig. 10A). As an example, COL2A1 (collagen type II alpha 1), associated with a strong cis-pQTL (rs2079295, P = 1 × 10⁻⁴⁰¹), maintained its stability (temporal SCC = 0.94 before vs. 0.89 after adjustment; Fig. 6B). These findings suggest that proteins affected by cis-acting genetic variation tend to be more temporally stable, even when accounting for genetic effects, while variable proteins generally lack cis-pQTL signals. Together with our previous observation that variable proteins tend to be intracellular and of lower abundance, the results indicate that their circulating levels are less genetically determined.

Fig. 6. Effect of pQTLs on temporal stability.

A Scatter plot of temporal SCC estimates before and after adjusting protein levels for conditionally independent genome-wide significant (linear regression P < 5 × 10⁻⁸) pQTLs. The original and adjusted temporal SCC values were highly correlated, with an SCC of 0.93. Proteins are colored by pQTL type (only cis, only trans, both, or none), with opacity indicating the significance of a protein’s strongest pQTL. B, C Examples of proteins initially classified as temporally stable; spaghetti plots show protein levels at the two study visits, before and after pQTL adjustment, in 100 randomly selected AGES participants. B COL2A1 remained temporally stable after adjustment (original SCC = 0.94; adjusted SCC = 0.89). Points are colored by participant’s genotype for the strongest pQTL for this protein (rs2079295, linear regression beta = −1.12, s.e. = 0.024, P = 1 × 10⁻⁴⁰¹). C ACP1 (original SCC = 0.85) initially showed genotype-based separation before adjustment but was reclassified as temporally variable after adjustment (SCC = 0.35), with reduced genotype separation and more apparent within-individual fluctuation. Points are colored by participant’s genotype for the strongest pQTL for this protein (rs79716074, linear regression beta = 1.26, s.e. = 0.021, P = 1 × 10⁻⁶⁰⁸).

While the absolute change in temporal stability following adjustment for genetic variants was not large (median 0.61 vs 0.59, respectively), the effects were widespread, as following adjustment, 75% of SOMAmers exhibited lower SCC values, indicating that they had more variable levels over time when genetic effects had been accounted for. There was a significant difference between original and adjusted temporal SCC across all pQTL groups (Supplementary Fig. 10B). ACP1 (acid phosphatase 1), originally stable (temporal SCC = 0.85) and associated with a strong cis-pQTL (rs79716074, P < 1 × 10⁻⁶⁰⁸), showed the most pronounced decrease in stability after adjustment (temporal SCC = 0.35) (Fig. 6C). Moreover, ACP1 exhibited high temporal variability within each genotype, suggesting that its apparent stability was largely genotype driven. The ten proteins with the largest decrease in temporal SCC (ranging from 0.39 to 0.50) were ACP1, CES3, TCN2, AHSG, UBE4A, ICAM1, MVK, GDF5, PILRA, and PRAME, of which five were reclassified from temporally stable to variable (Supplementary Data 19). Many of these proteins are influenced by strong genetic signals and have been implicated in various biological processes and disease mechanisms. For instance, ICAM1 (intracellular adhesion molecule 1), which has a key role in inflammation and immune cell adhesion has been implicated in various diseases such as autoimmune disease, sepsis, cancer, and cardiovascular diseases³⁹. PILRA (paired immunoglobulin-like type 2 receptor alpha) has also been proposed as a potential therapeutic target for Alzheimer’s disease⁴⁰ and atrial fibrillation⁴¹. Together, these results indicate that strong genetic effects on protein abundance can create the appearance of temporal stability, which underscores the importance of accounting for genetic drivers when interpreting longitudinal protein profiles in the context of disease.

Protein temporal stability in the context of disease

We have previously identified and replicated proteomic associations with aging-related diseases in AGES (Supplementary Data 2). Here, we examined whether proteins classified as temporally stable or variable were more likely to be associated with a range of prevalent and incident disease phenotypes and overall mortality in the AGES cohort (see “Methods”). We observed significant (FDR < 0.05) associations between baseline protein levels and all tested phenotypes, across both stable and variable proteins. For most diseases, a greater proportion of the associated proteins were temporally stable than variable (Fig. 7A; Supplementary Data 20). However, variable proteins were also linked to all phenotypes examined, with three related prevalent diseases (atrial fibrillation, stroke, and non-AD dementia) being associated with a significantly (two-sided Fisher’s exact test, FDR < 0.05) higher proportion of variable proteins (Fig. 7A; Supplementary Data 20).

Fig. 7. Temporal stability and disease associations.

A Number of proteins measured at baseline that are significantly associated with each phenotype (excluding proteins with medium temporal stability) in AGES, shown for both incident and prevalent disease outcomes. Bars are colored by protein temporal stability. An asterisk (*) above a phenotype indicates significant enrichment (two-sided Fisher’s exact test, FDR < 0.05) of either temporally stable or variable proteins associated with that phenotype. B Number of proteins with a significant interaction (linear regression, FDR < 0.05) of time of disease diagnosis on protein temporal stability. Data are stratified by time of diagnosis, temporal stability protein group, and direction of temporal stability change (positive or negative effect). C IL4 protein levels at AGES baseline and follow-up, stratified by AF diagnosis status: diagnosed before baseline, between baseline and follow-up, after follow-up, or never diagnosed. Regression slopes from the interaction model are shown for each AF group, and an asterisk (*) indicates a significant interaction in temporal stability change compared to the never-diagnosed group.

Given these associations, we next asked whether the presence or timing of disease could influence protein temporal stability. To do this, we assessed whether disease status modified temporal stability by fitting interaction models between baseline protein levels and disease status with follow-up protein levels as outcome. Disease status (timeline group) was defined as diagnosis: a) before the baseline visit, b) between baseline and follow-up, c) after follow-up, or d) never diagnosed (reference). We performed these analyses for seven diseases (Alzheimer’s disease [AD], atrial fibrillation [AF], coronary heart disease [CHD], congestive heart failure [CHF], non-AD dementia, myocardial infarction [MI], and stroke). Overall, 5164 of 7097 proteins (73%) showed significant (FDR < 0.05) interaction with disease status (FDR < 0.05; Supplementary Data 21), indicating that their temporal stability differed by disease. However, the proportions of temporally stable and variable proteins with significant interaction were similar (71% vs 72%).

Across all proteins, diseases, and disease status categories, 10,242 of 149,037 tests (6.9%) showed a significant (FDR < 0.05) interaction, indicating that disease status modifies protein temporal stability (Supplementary Data 21). Of these, 14% were associated with AD, 19% with AF, 12% with CHD, 14% for CHF, 12% for non-AD dementia, 15% for MI, and 13% for stroke. Among temporally stable proteins with significant interactions, nearly half (48%) were associated with diagnoses occurring between baseline and follow-up, suggesting that disease development between visits most frequently altered their stability. Moreover, 82% of these proteins showed reduced stability, reflected by a weaker baseline–follow-up association (negative interaction effect, Fig. 7B). In contrast, temporally variable proteins showed the highest proportion of significant interactions within individuals diagnosed before baseline (43%), indicating greater stability differences in those with pre-existing disease. Additionally, 72% of these proteins exhibited increased stability (positive interaction effect, Fig. 7B). Interestingly, high proportion of temporally variable proteins showed increased temporal stability in individuals diagnosed with AF (n_prot = 124, 22% of all variable proteins), and stroke (n_prot = 69, 12% of all variable proteins) before baseline (Fig. 7B). As an example, IL4 (Interleukin 4), a Th2 cytokine, showed significantly greater temporal stability in individuals with prevalent AF compared to those never diagnosed (regression slope = 0.74 vs 0.40, interaction FDR = 1.99 × 10⁻⁸, Fig. 7C). IL4 is known to promote fibroblast activation and collagen deposition in affected tissues, contributing to fibrosis in organs including the heart^42,43. Thus, our findings suggest that IL4 levels, normally variable over time, become more stable as AF develops and progresses.

Temporal stability of protein modules in aging and disease

Although temporally stable proteins were generally more likely to associate with disease phenotypes than variable proteins, many proteins are correlated or functionally linked, meaning that interpreting associations individually can give an incomplete picture⁴⁴. Taking into account the co-regulatory structure of the circulating proteome, we evaluated the temporal stability of proteins at the level of co-regulated protein modules. We applied weighted gene co-expression network analysis (WGCNA) using all 7288 proteins to define 25 co-expressed protein modules at baseline and 23 modules at the follow-up visit in the 3093 individuals with longitudinal measurements (Supplementary Data 22 & 23). The co-regulatory modules were enriched for different biological functions (Supplementary Data 24). We compared the baseline modules to the previously constructed co-regulatory network from the AGES cohort using the 5k version of the SomaScan assay (v3) in the full AGES baseline sample (n = 5457)⁷. The overall structure was similar, although some distinct clustering patterns were observed in this subset of individuals (Supplementary Fig. 11).

Integrating information on protein temporal stability, we identified 11 baseline modules enriched for stable proteins, 10 for variable proteins, and 4 for neither (Wilcoxon rank-sum test, two-sided, P < 0.05; Supplementary Data 24). An example of a module enriched for temporally stable proteins is M16 (n_prot = 192, P = 1.21 × 10⁻²²), which is composed of proteins localized to mitochondrial outer membranes, and the protein with the highest intramodular connectivity (hub protein, kWithin = 26.36) was GEM (GTP-binding protein), a small RGK family GTPase involved in calcium channel regulation and intracellular signaling⁴⁵. In contrast, M8 (n_prot = 300) is an example of a module enriched for temporally variable proteins (P = 3.69 × 10⁻³⁷), involved in coagulation cascades, where the hub protein was BTK (Bruton tyrosine kinase; kWithin = 27.23), a non-receptor kinase expressed in pre-B cells and B-lymphocytes⁴⁶ (Fig. 8A; Supplementary Data 24).

Fig. 8. Temporal stability of protein modules in association with disease.

A Examples of baseline modules enriched for each temporal stability (TS) group. Proteins in module M16 (n = 192) had significantly higher median temporal SCC estimates compared to other proteins (n = 7096; Wilcoxon rank-sum test, two-sided, P = 1.21 × 10⁻²²). Proteins in module M8 (n = 300) had lower median temporal SCC estimates compared to other proteins (n = 6988; Wilcoxon rank-sum test, two-sided, P = 3.69 × 10⁻³⁷). Boxplots indicate median value, 25^th and 75^th percentiles. Whiskers extend to the smallest/largest value of no further than 1.5 $\times$ interquartile range. Outliers are not shown. B Sankey plot showing how proteins are grouped into modules at each time point. Modules enriched for variable proteins are colored green, stable-enriched modules are colored blue, and modules without significant enrichment are colored gray. C Disease association enrichment within each baseline module. Modules are labeled on the x-axis and ordered by median temporal SCC. The y-axis lists each phenotype tested for associations with baseline protein levels. Orange dots indicate significant enrichment for a phenotype within a module, with dot size reflecting the strength of enrichment (odds ratio). Modules are highlighted vertically, with green indicating temporally variable-enriched and blue indicating temporally stable-enriched. D–G Temporal SCC is shown for proteins from three modules (M16, M2, M8) that had baseline levels associated with disease and differed in temporal stability depending on disease stage. There was a significant difference in median temporal SCC (empirical P < 0.01 from 10,000 permutations, see “Methods”) for individuals diagnosed between baseline and follow-up visits, compared to the other groups: those diagnosed before baseline, those diagnosed after follow-up, or those who were never diagnosed. Two modules were temporally stable-enriched: M16 (D–E) and M2 (F), both of which showed reduced temporal SCC in the between-visits disease groups for stroke and AF, or MI, respectively. The temporally variable-enriched module M8 (G) showed the opposite pattern, with increased temporal SCC in individuals diagnosed with dementia between visits.

We next explored the longitudinal behavior of co-regulatory patterns by examining the flow of proteins between baseline and follow-up modules (Fig. 8B). Overall, 5738 (79%) of the 7288 proteins remained within their dominant module group at follow-up, demonstrating a consistent co-regulatory pattern over time. Most modules retained a consistent protein composition over time, with only a few exhibiting reorganizations with proteins shifting to different modules at follow-up (Supplementary Data 25). There was no statistically significant difference in protein retention by temporal stability enrichment (stable-enriched modules, 78.6%; variable-enriched modules, 73.9%; no enrichment, 83.1%; Kruskal-Wallis χ² = 0.61; df = 2, P = 0.74). Thus, our results indicate that temporal stability in circulating protein abundance does not translate into the stability of inter-protein relationships.

In the preceding analyses, we investigated the relationships between protein temporal stability and disease status at the individual protein level. Accounting for the co-regulatory structure of the proteome, we next examined whether disease-associations (FDR < 0.05) were enriched within the co-regulatory modules at baseline (two-sided Fisher’s exact test, see “Methods”). Both stable- and variable-enriched baseline modules were enriched for disease associations (Fig. 8C; Supplementary Data 24) and comparing disease scores across modules stratified by temporal stability enrichment (see “Methods”) revealed no significant difference (Kruskal-Wallis χ² = 1.31, df = 2, P = 0.52; Supplementary Fig. 12). Despite this, we identified several stable modules with particularly high disease scores, including M15 (n_prot = 26, fatty acid and lipid metabolism), M13 (n_prot = 72, RNA metabolism), and M2 (n_prot = 781, cell adhesion and extracellular interactions) (Fig. 8C; Supplementary Fig. 12).

Exploring how disease status affected temporal stability at the module-level, we identified three modules containing disease-associated (FDR < 0.05) proteins with a significant (empirical P < 0.01) difference in median temporal SCC in individuals diagnosed between baseline and follow-up, compared to each of the other three groups (diagnosed before baseline, after follow-up, or never diagnosed) (Fig. 8D–G; Supplementary Data 26; see “Methods”). Within module M16, stable proteins associated with prevalent AF or stroke showed lower temporal SCC in individuals diagnosed between baseline and follow-up (median = 0.66 and 0.62, respectively) compared to others (median ≥ 0.72 and 0.8, respectively) (Fig. 8D–E). The proteins contributing most strongly to these differences were ARSB, IL17A | IL17F, and RLBP1 for stroke, and NPFF, PNPO, and FLJ44635 for AF. A similar pattern of reduced temporal SCC in individuals diagnosed with incident MI between visits was observed in module M2 (Fig. 8F). The proteins with the largest differences in this module were GDF15, TNFRSF1B, and RNASE1. Finally, variable proteins within module M8 associated with prevalent non-AD dementia showed a reversed pattern, as individuals diagnosed between visits had higher temporal SCC (median = 0.44) compared to others (median $\le$ 0.3) (Fig. 8G). The proteins most strongly driving this reversed pattern were YARS1, TIGAR, and CDC42. Thus, these findings illustrate how the onset of a disease state can influence the dynamics of co-regulated protein structures.

Discussion

In this study, we investigated the longitudinal stability of serum proteins over a five-year period in older individuals to characterize their temporal dynamics. Our findings reveal wide variability in the temporal stability of individual proteins and that temporal stability is largely independent of both the magnitude of change in protein levels between visits. We replicate protein temporal stability patterns across platforms in an independent cohort and demonstrate clear biological differences between proteins that tend to remain stable in circulation over time and those that vary. We also find that both genetic factors and disease state can influence protein temporal stability, which should be considered in disease biomarker research.

We found that temporally stable proteins in the circulation were more likely to be extracellular, secreted, and tissue- and organ-enriched (especially from adipose, immune, liver, pancreatic, and prostate tissues), and to exhibit more detected genetic associations. Somewhat surprisingly, we observed that temporally variable proteins generally had biological properties consistent with housekeeping genes, which are known to exhibit low variability in expression across tissues and conditions⁴⁷. Our comparison with external transcriptomic data, however, revealed that variable proteins were more likely to have low gene expression variance, whereas stable proteins showed the opposite. This pattern suggests that the long-term temporal stability of proteins in circulation is mainly driven by post-transcriptional regulation, release dynamics, or clearance, rather than reflecting transcriptomic regulation. At the levels of co-regulatory protein modules, we found that ~80% of the proteome remained in the same module over the five-year study period, suggesting that the co-regulatory structure of the circulating proteome is largely stable over time. Longitudinal module retention was not related to the temporal stability of individual proteins, demonstrating that temporally variable proteins can co-vary and maintain stable inter-protein relationships. Overall, our findings provide important insights into the pathways and functions of circulating proteins that differ in relation to their behavior over time.

Generally, we find that temporal stability increases with age. The proteins showing consistent effects across the AGES and CHS cohorts included key inflammatory regulators (TNFRSF1B, SIGLEC5, NCR1, ESAM) and antioxidant enzymes (SOD1, CAT), suggesting that these proteins require more consistent regulation during aging. Overall, proteins showed more temporal variability in females, even when levels did not differ between sexes, demonstrating sex-specific protein dynamics in older adults.

Previous studies have described longitudinally stable proteomic fingerprints of individuals, driven by the strong genetic influence on protein levels^23,48. Consistent with these findings, when examining individual protein patterns, our study shows that genetic variability significantly influences the temporal stability of serum proteins. Adjusting for pQTLs decreased temporal stability for 75% of proteins and, notably, some proteins initially classified as stable were even reclassified as variable, suggesting that genetic factors contribute substantially to their observed stability over time. Such genetic effects are particularly important to consider for biomarker candidates, as they both mask population-level disease associations and person-specific deviations from their baseline over time.

Circulating proteins generally need low biological variability within and between individuals to be useful as clinical biomarkers⁴⁹. Although we did not find a consistent relationship between protein intra- and inter-individual variability, baseline levels of temporally stable proteins were generally more likely to be associated with disease outcomes than of variable proteins. This information, together with the stable proteins’ extracellular nature and tissue-specific expression, makes them valuable candidates for drug targets and disease biomarkers⁵⁰. Investigating how disease manifestation influenced temporal stability, we observed that proteins that are generally stable over time exhibited more temporally variable patterns in the context of disease manifestation for a range of diseases. Specifically, individuals diagnosed between visits showed markedly greater temporal variability in protein levels compared to others. Interestingly, we also observed that the pattern was reversed for temporally variable proteins, with their levels becoming more stable, particularly after the onset of disease. Thus, the temporal stability of specific proteins can either increase or decrease during disease development, and further studies are needed to elucidate the underlying mechanisms for individual protein candidates.

Examining disease state effects on temporal stability while accounting for the co-regulatory structure of circulating proteins revealed consistent patterns. Temporally stable proteins in module M16 showed decreased stability at the onset of AF and stroke, and were enriched for mitochondrial outer membrane proteins, a cellular component known to regulate apoptotic signaling pathways⁵¹. Mitochondrial dysfunction is also a well-established contributor to the development of cardiovascular diseases⁵². One protein showing a particularly large change in temporal SCC in individuals who developed stroke between visits was IL17A | IL17F, previously linked to stroke-related neuroinflammation⁵³. Another example is module M2, which included temporally stable proteins whose stability decreased in individuals with MI events. This module contained established markers of cardiovascular stress and inflammation, such as GDF15, TNFRSF1B, and RNASE1, and was enriched for proteins involved in cell adhesion and extracellular interactions. This pattern, in which temporally stable proteins exhibit decreased temporal stability across visits, suggests dysregulation of processes related to disease onset or progression.

Interestingly, module M8 included temporally variable proteins with a reversed pattern compared to the other examples, that is, exhibiting greater stability in individuals diagnosed with non-AD dementia between visits. This suggests that proteins in M8 are generally temporally variable under non-disease conditions. However, during the window surrounding dementia diagnosis, their levels become more stable. The top proteins showing the greatest stability differences are involved in neuronal stress response (TIGAR, YARS1, CDC42, TIA1), and are thus promising candidates for further mechanistic investigation, as these protein dynamics could provide novel insights into the process driving dementia progression to a clinically detectable stage.

The major strengths of our study include a large sample size, broad proteome coverage, replication and validation of findings in an independent cohort, and evaluation of long-term stability in individuals at an age where common diseases are likely to manifest. However, some limitations must be acknowledged. One is the availability of only two measurements per individual. Longitudinal proteomics studies that include multiple time points enable a more detailed analysis of protein patterns over time and the effects of aging on circulating protein levels. With more than two visits, alternative methods, such as the intraclass correlation coefficient (ICC)⁵⁴ could be used to determine temporal stability. For a more accurate validation of technical variability, the inclusion of duplicate samples from individuals in the AGES cohort would have been ideal. However, in the absence of such duplicates, we used reference values from the ARIC study to gain an approximate understanding of the SOMAmers‘ technical variability. It is important to note that technical variability may differ between the AGES, ARIC, and CHS cohorts due to differences in sample handling or processing protocols⁵⁵. Finally, as our analyses were limited to individuals with protein measurements at both visits, there is potential for selection bias. Participants who did not return for follow-up may have been in poorer health or at greater risk of disease, potentially leading to an underestimation of temporal variability or disease associations in our sample.

In conclusion, we have shown that multiple factors influence the long-term temporal stability of circulating proteins. Temporally stable and variable proteins appear to represent distinct sets of proteins with different biological functions, primarily reflecting extracellular and intracellular processes. We also demonstrated that protein temporal stability appears to reflect the mechanisms underlying proteins’ presence in circulation rather than their transcriptomic regulation. Furthermore, temporal stability of proteins can differ with age and between sexes. The most widespread effects of temporal stability were genetic effects and disease stage, providing valuable information for biomarker discovery and precision medicine.

Methods

Study cohort

The AGES-Reykjavik study was approved by the National Bioethics Committee in Iceland (approval number VSN-00-063), the National Institute on Aging Intramural Institutional Review Board, and the Data Protection Authority in Iceland. All participants provided informed consent before study enrollment. Study participants did not receive any compensation for their participation. This study was conducted by the Icelandic Heart Association (IHA) and is a population-based, single-center, prospective cohort study initiated in 2002 and with baseline examinations completed in 2006⁵⁶. The AGES-cohort comprises 5764 Icelandic individuals aged 66 to 93, with a mean age of 76. Of these, 3316 participants returned for a follow-up visit ~five years later (2007–2011). Genotypic and proteomic data at two time points were available for 3093 individuals. No statistical method was used to predetermine sample size. Information about multiple phenotypes and traits for AGES participants at both visits was available⁵⁶, along with linked medical records. Supplementary Data 2 provides a summary of the phenotypes, abbreviations, regression models, and covariates used.

Protein data

Blood samples were collected at both AGES-Reykjavik visits after an overnight fast. Serum was prepared using a standardized protocol, stored in 0.5 ml aliquots at −80 °C, and serum samples that had not been thawed previously were used for protein measurements. All AGES samples are stored in the Icelandic Heart Association Biobank. Serum protein levels were measured at both visits in the AGES study using the SomaScan (Slow Offrate Modified Aptamer Scan, v4.1) platform, analyzing 7523 proteins⁵⁷. To avoid batch or processing time biases, the order of sample collection and processing for protein measurements was randomized, and all samples were run as a single batch and blinded to any phenotypic outcomes at SomaLogic Inc. (Boulder, CO, US). The SOMAmers had a median intra- and inter-assay coefficient of variation (CV)  <  4%. Hybridization controls were used to correct systematic detection variability, with calibration samples measured across three dilution sets: 20% (1:5), 0.5% (1:200), and 0.005% (1:20,000), ensuring that fluorescence reflected protein concentration quantitatively. The adaptive normalization by maximum likelihood (ANML) method was employed to normalize Quality Control (QC) replicates and samples using point and variance estimations from a healthy external reference population (n = 1000). Three outlier samples were excluded based on a principal component analysis. Consistent target specificity of aptamers was indicated by direct (through mass spectrometry) and/or indirect validation⁷.

A total of 6385 serum proteins (referred to by their gene symbols) were targeted by 7288 SOMAmer reagents using the high-throughput SomaScan proteomic platform (v4.1), after excluding 281 SOMAmers derived from mouse-human chimeras. To ensure comparability across proteins and mitigate potential skewness in the data distribution, the Box-Cox transformation was applied, and values were standardized to a mean of 0 and a standard deviation of 1.

Validation cohort (CHS)

For replication and validation, we analyzed data from the Cardiovascular Health Study²⁹ (CHS, mean age = 72 years). CHS is a population-based longitudinal study of 5888 community-dwelling older people recruited from four United States communities: Sacramento County, CA; Washington County, MD; Forsyth County, NC; and Allegheny County, PA. Between 1989 and 1990, 5201 participants aged 65 years or older were recruited from Medicare eligibility lists, with an additional 687 predominantly African American participants enrolled in 1992-1993. EDTA-plasma was collected from fasting blood samples in 1992-1993 and 1996-1997 and stored at –70 to –80 °C until proteomic profiling was performed. Each participant gave informed consent, and each center underwent institutional review board approval.

In 2020, proteomic profiling was performed on stored samples from study year 5 (1992-1993) for 3171 participants using a SomaScan panel (5k SomaScan v4.0 assay). In 2021, proteomic profiling was performed in an additional 485 participants in stored samples from study year 5 (1992-1993) using the SomaScan v4.1. In 2024, using the Olink Explore HT platform, plasma proteins were measured in stored samples from CHS year 5 (1992-1993) in 4037 participants and from year 9 (1996-1997) in 2928 participants. Of these, 2770 participants had Olink measurements at both time points. There were 3095 CHS participants whose samples were measured with both SomaScan and Olink platforms. Across platforms, 3402 proteins overlapped between Olink and SomaScan, of which 990 were unique to the SomaScan 7k platform. To enable validation of the AGES SomaScan analyses, all Olink protein levels at both time points in CHS were scaled to a mean of 0 and a standard deviation of 1.

Genetic data and protein quantitative trait loci (pQTLs)

AGES participants (n = 5661) were genotyped using the Illumina Infinium Global Screening Array (GSA, n = 2710) and Illumina hu370CNV array (n = 2951), as previously described^4,38. Genotype imputation was performed using the TOPMed reference panel⁵⁸, yielding 11,919,579 variants after quality control and merging data from both platforms. We previously conducted a protein GWAS of the 5k SomaScan data (v3)⁴, and an updated analysis based on the 7k dataset (v4.1) was used in this study³⁸. Each of the 7569 SOMAmers measured in AGES was tested for association with 11,919,579 variants in a linear regression using PLINK 2.0, adjusting for sex, age, 10 genetic principal components, and genotyping platform. Stepwise conditional analysis was then applied with GCTA-COJO⁵⁹ to identify independent protein quantitative trait loci (pQTLs) for each SOMAmer. The variant with the lowest P at each step was designated as the lead variant. Variants in strong linkage disequilibrium (LD; r² > 0.9) with the lead variant were excluded from joint analysis to avoid multicollinearity. pQTLs with an independent lead variant within ±500 kb window of the protein-coding gene were defined as cis-acting; otherwise, they were classified as trans-acting.

Statistics and reproducibility

Statistical analyses were performed in RStudio software (R version 4.3.2⁶⁰), unless otherwise noted. All protein values were included for Spearman correlation analyses, while extreme outlier values (>4.3 SD from mean)⁹ were excluded from linear regression models. The temporal stability of each protein between baseline and follow-up visits was assessed using the Spearman correlation coefficient (SCC), hereafter referred to as the temporal SCC. Proteins with SCC higher than 0.75 were defined as temporally stable (n_prot = 729), whereas proteins with SCC lower than 0.40 were defined as temporally variable (n_prot = 681) based on suggestive guidelines for the interpretation of SCC²⁵ and as they corresponded closely to the 90^th (0.754) and 10^th (0.406) percentiles of the distribution. The proteins that fell between 0.4 and 0.75 in temporal SCC were defined as having medium temporal stability.

In the CHS cohort, platform correlations of 3402 proteins between SomaScan and Olink were calculated using the participants with both assays available at year 5. Proteins with low platform correlations (SCC < 0.4) were removed prior to validation and replication analyses. When a single SOMAmer matched multiple Olink IDs, the ID with the highest SCC was selected. This filtering yielded 1301 unique proteins (SOMAmers) for validation. Temporal stability of plasma proteins in CHS (Olink) was then assessed in the 2770 participants with longitudinal data, using the same pipeline and thresholds applied to SomaScan data in AGES. Finally, temporal SCC estimates across datasets were compared using SCC.

To assess whether temporal SCC provided different insights into longitudinal variability compared to the absolute difference in protein levels between time points ($\mathrm{\Delta },$ delta), the median delta value per protein across individuals was computed, and its relationship with temporal SCC estimates evaluated using the SCC. To investigate the relationship between inter- and intra-individual variability in protein levels, the interquartile range (IQR) for each protein at each time point was computed and compared with temporal SCC estimates.

We tested whether relative protein abundance (raw protein measurements in RFU units, without transformation) differed between the temporal SCC groups. To account for differences in dilution, the median raw protein value for each protein at the baseline visit was computed and adjusted using the corresponding dilution factor. Thus, proteins in the 20% dilution set were multiplied by 1/20, while proteins in the 0.5% dilution set were multiplied by 1/0.5. To assess the relationship between temporal SCC and abundance, the distributions of median dilution-corrected RFU multiplied by the dilution factor for each protein were compared between temporal stability groups using a pairwise Wilcoxon test (two-sided). The relative protein abundance was additionally compared with external concentration values from the Human Protein Atlas⁶¹ (v23.0), which provided immunoassay-based protein concentrations (pg/mL) for 569 of the 7288 proteins in our study, using SCC.

For the assessment of technical variability of SOMAmers, external reference values from the ARIC study were used to analyze duplicate samples from 102 individuals using the same 7k SomaScan platform (v4.1)³⁰. This study estimated interplate stability using three different methods. For this study, the percentile variation (PV) values from the ARIC study were chosen because they demonstrated the least sensitivity to outliers and did not require the assumption of a normal distribution. Proteins with PV values more than two standard deviations (SD) above the mean were defined as having substantial technical variability. In total, 191 proteins exceeded this threshold and were classified as technically variable. These proteins were excluded from downstream analyses to reduce the influence of technical variability on the biological interpretation of temporal stability.

Pathways and functional enrichment analyses were conducted using g:Profiler⁶² (g:GOST module) to identify potential functional differences between stable and variable proteins. A list of proteins in each temporal SCC protein group was used as an input, with the 7097 measured proteins defined as technically stable, serving as a custom background. The significance threshold was set at the Benjamini-Hochberg false discovery rate⁶³ (FDR) < 0.05. Data sources included Gene Ontology⁶⁴ (GO), KEGG⁶⁵, Reactome⁶⁶, and WikiPathways⁶⁷. For proteins with multiple GO annotations under the same Ensembl ID, annotations with the highest number of associated GO terms were prioritized. To identify key differences between stable and variable proteins, driver GO terms were extracted and compared between the groups. Additionally, a multi-query analysis was performed for the AGES-defined temporal stability groups, in which both lists of stable and variable proteins were analyzed simultaneously using the complete set of 7097 proteins as a background. By conducting this simultaneous analysis, the significance of enriched terms between stable and variable protein groups was directly compared, providing insights into their functional differences. To validate the findings, the same enrichment analyses were performed in the independent CHS validation cohort using proteins classified as temporally stable or variable, independent of AGES results. These analyses were restricted to proteins that showed at least moderate correlation between the SomaScan and Olink platforms and were previously defined as technically stable in SomaScan. This resulted in 1301 proteins used as background, of which 243 were stable, and 122 were variable in CHS.

Protein class enrichment in AGES was examined using external data from the Human Protein Atlas⁶¹ (v23.0), which included 5894 proteins overlapping with the SomaScan 7k protein dataset. Two-sided Fisher’s exact test⁶⁸ was applied to identify protein classes that were significantly overrepresented in either the temporally stable or variable protein groups. Multiple testing was corrected using FDR (FDR < 0.05). Additionally, using the same type of tests, enrichment for tissue-enriched genes⁶⁹ and protein expression⁷⁰ was assessed between temporal SCC groups.

Tissue- and organ-specificity between temporal SCC groups in AGES was assessed using external RNA-seq annotation data from GTEx (v10)³¹. This dataset included expression data for 6963 SOMAmers corresponding to 6120 unique Entrez gene symbols after filtering out proteins with high technical variability. Organ specificity for each SOMAmer was defined using a tissue-to-organ mapping based on the previously published approach by Oh et al.⁷¹. For each SOMAmer, the top tissue and top organ were defined as those with the highest median gene expression. Two-fold enrichment was defined as cases in which expression in the top tissue or organ was at least 2-fold higher than in the second-highest tissue or organ. Four-fold enrichment was defined analogously using a four-fold expression difference. A two-sided Fisher’s exact test was used to determine whether specific organs or tissues were significantly enriched among temporal SCC protein groups, with multiple testing corrections using FDR.

Potential differences in protein interactions between the stable and variable protein groups in AGES (a total of 1403 proteins excluding the medium temporal stability group) were explored using both their hub status in the InWeb protein-protein interaction (PPI) network³² (data available for 568 proteins) and the serum protein co-regulatory network⁷ (data available for 788 proteins). This was performed using a two-sided Wilcoxon test. Additionally, LoF intolerance was assessed using pLI scores⁷² for the protein-encoding gene (data available for 1138 proteins). Genes were classified as LoF intolerant (pLI $\ge$ 0.9) or LoF tolerant (pLI <0.9)⁷², and enrichment in temporally stable and temporally variable protein groups was evaluated using a two-sided Fisher’s exact test.

Gene expression variability was investigated to determine whether it correlated with protein temporal stability. A previous study analyzed 57 publicly available RNA-sequencing datasets to examine variance in gene expression across different tissues and datasets³³. In this study, gene expression variance was quantified as the standard deviation (SD) of gene expression levels across individuals for each gene, and each gene was ranked by its SD. SD-based gene expression ranks were available for 4367 SOMAmers measured in AGES, enabling a comparison of temporal SCC groups using a two-sided Wilcoxon rank-sum test.

To assess whether temporal stability differed by age or sex, we examined the association between baseline and follow-up protein levels using linear regression models. For each protein, follow-up levels were modeled as a function of baseline levels, age, and sex. Effect modification by age and sex was evaluated by including interaction terms between baseline protein level and age or sex, respectively, in otherwise identical models. For comparison, associations between protein levels at each time point and age and sex were also tested separately. To validate the findings, the same analyses were performed on the Olink Explore HT plasma protein dataset from the CHS cohort.

Protein quantitative trait loci (pQTL)s for the SomaScan 7k measurements in AGES (see Genetic data and protein quantitative trait loci (pQTL)) were used to assess the influence of genetic variants on the temporal SCC. A linear regression model was used to adjust each of the 7288 protein levels for their respective conditionally independent pQTLs. Specifically, for each protein, all conditionally independent genetic variants with genome-wide significant protein associations (P < 5 × 10⁻⁸) for that protein were included as covariates in the regression model. The residuals from this model were extracted and used as the adjusted protein values. For the remaining 716 proteins (SOMAmers) without genome-wide significant pQTL associations, their original values were retained to ensure a balanced comparison across all proteins. Temporal SCC for the adjusted protein values was then recalculated using the same approach described above and compared with the original temporal SCC estimates.

The association between temporal stability and age-related diseases was assessed using three different models. Protein levels at baseline were tested for association with each phenotype using the full cohort, adjusted for age and sex. For prevalent phenotypes (diagnosed before or at the baseline visit), a logistic regression model was used. A Cox regression model was used for incident phenotypes (diagnosed after baseline-visit) to include the time between baseline and diagnosis, except for type 2 diabetes and chronic kidney disease, which were defined based on data from the follow-up visit and not time-to-event data (see Supplementary Data 2 for details), and thus analyzed using logistic regression. The glm() function was applied for the logistic regression analysis, and for Cox regression, the coxph() function from the survival package⁷³. Multiple comparisons were controlled using FDR correction, where proteins with FDR < 0.05 were considered significantly associated with their corresponding phenotype.

Seven phenotypes (myocardial infarction [MI], coronary heart disease [CHD], congestive heart failure [CHF], atrial fibrillation [AF], Alzheimer’s disease [AD], non-AD dementia, and stroke) were examined to assess whether disease status modifies temporal stability. For each phenotype, participants were categorized into one of the four disease diagnosis timeline groups:

• Before baseline: diagnosed before baseline visit

• Between baseline and follow-up: diagnosed between baseline and follow-up visit

• After follow-up: diagnosed after follow-up visit

• Never: Never diagnosed

The disease diagnosis timeline groups were then used as an interaction term with the baseline protein measurement in a linear regression model, using the follow-up protein measurement as the outcome. The lm() function was used for linear regression analyses, and multiple testing corrections were controlled using the FDR.

Weighted gene co-expression network analysis (WGCNA) was performed on protein expression levels separately for baseline and follow-up visits using the WGCNA R package⁷⁴. To enable comparison with previously generated modules from the 5k protein dataset⁷, we set the soft-thresholding power (β) to 5 for both baseline and follow-up analyses. Co-regulatory modules were identified using dynamic tree cutting. Pathway and functional enrichment analyses were conducted for each baseline module using g:Profiler⁶², with the complete set of the 7k protein platform used as the background.

Protein modules at baseline were tested for enrichment of temporal SCC groups using a two-sided Wilcoxon rank-sum test. In addition, enrichment for protein associations (FDR < 0.05) with 23 diseases or phenotypes (12 prevalent and 11 incident) was tested using a two-sided Fisher’s exact test, with FDR correction applied at a threshold of FDR < 0.05. Disease enrichment per module was summarized into a disease score by multiplying the number of significantly enriched diseases by the median odds ratio (OR) for those enrichments, thereby accounting for both the number and strength of disease associations within each module.

For proteins within each baseline module that were significantly associated with the respective phenotype, temporal SCCs were recalculated separately within each disease timeline group. A pairwise permutation test (10,000 permutations), permuting group labels, was used to determine whether the median temporal SCC among individuals diagnosed between visits differed significantly from other individuals. Only protein sets with an absolute difference ≥0.05 and at least 10 proteins were considered for testing.

Reporting summary

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

Author contributions

Conceptualization: H.K.I., Va.G. Formal Analysis: H.K.I., H.R.B., T.R.A., E.A.F., E.J., K.A. Resources: Vi.G., L.J.L., N.F., J.J.L., V.E, M.O., J.C.B. Data curation: H.K.I., T.A., T.R.A., Va.G. Writing – Original Draft Preparation: H.K.I., Va.G., Writing – Review and Editing: all authors, Supervision: Va.G., A.P., Vi.G.

Peer review

Peer review information

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

Funding

National Institute on Aging contracts N01-AG-12100 and HHSN271201200022C, the Icelandic Heart Association, and Althingi (the Icelandic Parliament) financed the AGES-Reykjavik study. IHA and Novartis have collaborated on proteomics research since 2012. H.K.I. and Va.G. disclose support for the research of this work from the Memorial fund of Helga Jonsdottir and Sigurlidi Kristjansson, the Icelandic Research Fund (2511008-051), and the Icelandic Center for Research (2412215-1101). Research reported in this publication was supported by the National Institute on Aging of the National Institutes of Health under Award Number R01AG065596 (Va.G., Vi.G., V.E., E.A.F., T.A., L.J.L.). The Cardiovascular Health Study (CHS) was supported by contracts HHSN268201200036C, HHSN268200800007C, HHSN268201800001C, N01HC55222, N01HC85079, N01HC85080, N01HC85081, N01HC85082, N01HC85083, N01HC85086, 75N92021D00006, and grants U01HL080295, U01HL130114, R01HL144483 and R01HL172803 from the National Heart, Lung, and Blood Institute (NHLBI), with additional contribution from the National Institute of Neurological Disorders and Stroke (NINDS). Additional support was provided by R01AG023629 from the National Institute on Aging (NIA) and by R01HL144483 from the National Institutes of Health. A full list of principal CHS investigators and institutions can be found at CHS-NHLBI.org. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Molecular data for the Trans-Omics in Precision Medicine (TOPMed) program was supported by the National Heart, Lung, and Blood Institute (NHLBI). Proteomics for “NHLBI TOPMed: Cardiovascular Health Study (phs001368.v4.p2 was performed at Baylor-UTHealth Metabolomics Center (HHSN268201600034I). Core support, including phenotype harmonization, data management, sample-identity QC, and general program coordination, was provided by the TOPMed Data Coordinating Center (R01HL-120393; U01HL-120393; contract HHSN268201800001I). We gratefully acknowledge the studies and participants who provided biological samples and data for TOPMed.

Data availability

Individual-level participant data from the AGES Reykjavik study are available through collaboration under a data usage agreement with the Icelandic Heart Association. Access to data is controlled due to participant privacy protections, ethical committee restrictions, and the terms of informed consent signed by participants. These restrictions do not allow for individual-level genotype data to be shared, although summary statistics can be made available. Data access requests should be directed to AGES_data_request@hjarta.is. The time it takes to respond to requests varies depending on their nature and the circumstances, although it rarely exceeds 14 working days. Approved data-use agreements restrict data use to the approved research purpose and prohibit attempts to identify participants or to share data with unauthorized third parties. Individual-level participant data from the Cardiovascular Health Study (CHS) are available under the CHS Data and Specimen Distribution Policy (https://chs-nhlbi.org/CHS_DistribPolicy), which outlines access conditions, application procedures, and data use restrictions. All summary data supporting the conclusions of the paper are presented in the main text and freely available as a supplement to this manuscript (Supplementary Information and Supplementary Data). The GTEx database (https://gtexportal.org/home/) was used to obtain expression data for tissue/organ enrichment analysis. Gene expression variance data were retrieved from Wolf et al.³³. Information on protein classes was obtained from the Human Protein Atlas (https://www.proteinatlas.org/humanproteome/proteinclasses).

Code availability

The code used in this study is accessible at the following repository: (https://github.com/huldakareningvars/Temporal_Stability).

Competing interests

N.F. and J.J.L. are employees and stockholders of Novartis. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at (10.1038/s41467-026-72957-w).