Activated GDF11/8 subforms predict cardiovascular events and mortality in humans

Abstract

Circulating Growth Differentiation Factors 11 and 8 (GDF11/8) exist in both latent and active forms, and it is unclear if specific forms can predict disease outcomes. Our data suggest that a dual-specific aptamer selectively binds GDF11/8 after prodomain activation. In 11,609 patients at risk for future cardiovascular events, low dual-specific aptamer-detected GDF11/8 levels strongly predicted adverse outcomes, including cardiovascular events (HR = 0.43, p = 9.1 × 10⁻⁶³) and all-cause mortality (HR = 0.33, p = 4.8 × 10⁻⁴⁰). Use of selective aptamers suggested that results observed with the dual-specific aptamer for cardiovascular and mortality risk replicated with a GDF8 aptamer although with a smaller effect size. In a second cohort of 4110 individuals (ARIC), low dual-specific aptamer-detected GDF11/8 levels also predicted increased 8 year dementia risk (HR = 0.66, p = 0.00148). Our findings reveal that activation of GDF11/8 may be a factor in future aging-related cardiovascular and cognitive decline.

The predictive value of blood levels of GDF11/8 proteins in humans for future disease outcomes has been unclear. Here, the authors show that low levels of specific activated GDF11/8 subforms are strongly associated with future cardiovascular events, mortality, and dementia risk.

Introduction

Heterochronic parabiotic experiments in mice reported that connecting the circulations of young (2 months old) and old mice (2 years old) appeared to reverse age-related changes in heart¹, brain², and skeletal muscle of aged mice³. Proteomic screening using aptamers as binding reagents identified growth differentiation factor 11 (GDF11), a transforming growth factor (TGF)-β superfamily member, as a circulating factor that might account for some of the observed beneficial effects in aged parabionts. It was initially reported that circulating levels of GDF11 decline with age¹. Those original studies found that administration of recombinant GDF11 to aged mice at doses that restored circulating GDF11 to its youthful levels recapitulated the beneficial effects observed in the parabiotic studies^{1, 2–3}. However, studies in certain other murine models did not show the same benefits as these initial findings⁴. Questions also arose about whether some of the effects attributed to GDF11 in aging animals should instead be ascribed to myostatin (GDF8)⁴.

GDF8 is a close structural homolog of GDF11, with 90% amino acid sequence identity shared in their mature active forms. Importantly, the two molecules could not be distinguished by most binding assays (including aptamers) used at the time of the earlier reports⁵. Lastly, both GDF11 and GDF8 can be found in multiple activity states, including immature forms in which a prodomain maintains these proteins in a latent state and a mature, activated form generated after the prodomain has been removed by enzymatic cleavage⁶. It has been unclear which of these forms of GDF11 and GDF8 were detected by the aptamer binding assays or by mass spectrometry⁴ in previously published studies (see below).

Disagreements regarding the effects of GDF11 and GDF8 across the various experimental models tested in mice might be resolved through carefully conducted human studies. Yet, despite years of effort, the roles of GDF11 and GDF8 in human cardiovascular disease remain unclear. Mass spectrometry studies of human blood in which all forms of GDF11 or GDF8 are measured, irrespective of their activity state, have not shown statistically significant correlations with the risk of cardiovascular disease⁷ or dementia⁸. However, long-term observational data from two human cohorts—each including almost 1000 patients with stable coronary heart disease—indicate that lower blood levels of an aggregate measure of GDF11 and GDF8 (hereafter denoted GDF11/8), detected using a dual-specific aptamer, predict a high risk of adverse cardiovascular outcomes and mortality over the ensuing 5−8 years⁹. In that study, levels of GDF11/8 detected in blood were lower in older participants. Importantly, these GDF11/8 measurements were made with a dual-specific aptamer (2765-4, as designated by Somalogic, Inc. (Boulder, CO)) that does not distinguish between GDF11 and GDF8. We speculated that the discrepant conclusions that have been reached using different approaches to measure GDF11/8 in other contexts could be harmonized if the dual-specific aptamer 2765-4 recognizes specifically those forms of GDF11 and GDF8 that relate to clinical outcomes such as heart failure, myocardial infarction, stroke, and death¹⁰. Such a situation would explain why changes in levels of the forms of GDF11 and GDF8 detected by 2765-4 would correlate with health outcomes, whereas changes in total levels of GDF11 and GDF8 (including all forms of these proteins) would not.

GDF11 and GDF8 are secreted as latent precursors cleaved by a Furin protease, separating the large N-terminal prodomain from the C-terminal mature signaling domain¹¹. Unlike most TGFβ superfamily proteins, GDF11 and GDF8 mature ligands are tightly bound to their prodomains, remaining inactive until tolloid proteases (TLDs) in the extracellular space cleave the prodomain. TLD cleavage leads to a triggered state allowing future release of mature ligands from prodomains^6,12,13. Here we show that the dual-specific aptamer 2765-4, which predicted the risk of adverse cardiovascular events in patients with stable coronary heart disease, recognizes GDF11 and GDF8 after, but not before, cleavage of the prodomain. This biochemical discovery led us to use this unique dual-specific GDF11/8 aptamer 2765-4 to study major cardiovascular and mortality outcomes in a broader group of individuals, a meta-cohort of 6 pooled studies with a wide range of cardiovascular risk that is more representative of the general population than the cohort of patients with stable angina investigated previously⁹; this meta-cohort is also six-fold larger than previously studied cohorts⁹. We also investigated an additional cohort from the ARIC study for prediction by GDF11/8 of the risk of incident dementia. To gain insight into whether the association of GDF11/8 with clinical outcomes reflected GDF8 and/or GDF11, we used aptamers specific to these two ligands. These findings reveal that it is the decreased presence of activated GDF11/8, specifically, that confers heightened risk of adverse clinical outcomes.

Results

Aptamer 2765-4 only binds a portion of GDF11/8 in human serum

Given the conflicting predictive findings from different GDF11 assays, we hypothesized that aptamer 2765-4 against GDF11/8 would not bind to some forms of GDF11/8 ligands, whereas liquid chromatography with tandem mass spectrometry (LC-MS/MS) would be expected to detect all forms of these ligands in human serum, as the peptide fragments measured are common to all forms. We quantified GDF11 and GDF8 levels, individually, in healthy human serum by LC-MS/MS, both before and after ligand pull-down by the dual-specific aptamer 2765-4. The amount of total ligand in the serum measured by LC-MS/MS is shown in Fig. 1A-C. The amount of ligand bound to aptamer 2765-4 in pull-down assays was likewise measured by LC-MS/MS, shown in Fig. 1D-F. Supporting our hypothesis, aptamer 2765-4 captured less than half of the total ligand measured by LC-MS/MS (Fig. 1G-I). With the very small sample size in these experiments (n=10 male and 10 female samples per age group), there was no relationship between total GDF11 and GDF8 ligands with age.

[See PDF for image]

Fig. 1

Aptamer only pulls-down a fraction of Growth differentiation factor 11/8 (GDF11/8) from healthy human serum.

A−C The amount of GDF11/8 ligands in healthy human serum samples, measured by LC-MS/MS. D−F The amount of GDF11/8 ligands pulled-down by the aptamer from the same human serum samples, measured by LC-MS/MS. G−I the ratio of the amount pulled-down aptamer to the LC-MS/MS measure. Young: 18−25 years old, Middle Aged: 40−60 years old, and Aged: older than 60 years old. Two-way ANOVA with Turkey’s multiple comparison test, n = 10 in each group. Statistical analysis: two-way ANOVA (two-sided) with Tukey’s post-hoc test for multiple comparisons. Data are presented as mean ± SEM. n = 10 biological replicates per group. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. LC-MS/MS = immunoplexed liquid chromatography with tandem mass spectrometry.

Dual-specific aptamer 2765-4 does not bind to inactive latent GDF11/8

Because dual-specific aptamer 2765-4 only bound to a minority of GDF11/8 ligands in human serum, we hypothesized that this aptamer bound to some forms of GDF11/8 and not to other forms. To determine which states of the ligands were recognized by the aptamer 2765-4, we expressed and purified different forms of latent GDF11 and GDF8 that are thought to exist in vivo⁶. Like other TGFβ proteins, GDF11 and GDF8 are made as inactive precursors (hereafter referred to as unprocessed GDF11/8) (Fig. 2A). These precursors are then processed by furin proteases to form an inactive latent complex⁶. When the prodomain is cleaved by extracellular tolloid proteases, the latent complex moves into the triggered molecular state that prepares the cleaved prodomain to release the mature ligand for signaling¹⁴. There are several known proteins that can bind to the mature ligands after their release from their prodomains, such as follistatin (FS), follistatin-like 3 (FSTL3), WAP, Kazal, immunoglobulin, Kunitz and NTR domain-containing protein 2 (WFIKKN2), and WFIKKN1^{6,11,15, 16–17}. We tested whether the dual-specific aptamer 2765-4, the GDF8 specific aptamer, or the GDF11 specific aptamer captured latent GDF11/8 (Fig. 2) and the mature ligands with and without inhibitors (Fig. 3). Since the wild-type latent GDF11 is extremely difficult to express and purify for technical reasons¹⁸, we used a chimeric GDF11/8 construct as a surrogate for latent GDF11. The construct is composed of the GDF8 prodomain and the GDF11 mature domain (referred to as GDF8P11M). The different signaling activities of the latent and acid-activated recombinant proteins were verified by luciferase assay (Fig. S1).

[See PDF for image]

Fig. 2

The dual-specific aptamer did not recognize the inactive GDF11/8 complex.

A Unprocessed GDF8/11 (inactive) undergoes Furin cleavage to form the latent complex (inactive). TLD processing converts the latent complex into the triggered state (active), and dissociation of the remaining prodomain yields an unbound mature ligand (active). The binding of an antagonist renders the ligand inactive. B, C Anti-GDF8 prodomain (B), and anti-GDF11/8 mature ligand (C) western blot of the latent complex (Latent GDF8 and Latent GDF8P11M; GDF11 mature ligand with GDF8 prodomain), latent complex treated with BMP1 (Latent GDF8 / BMP1 and Latent GDF8P11M / BMP1), and latent complex incubated with the same digestion conditions but without BMP1 (Latent GDF8 / BMP1 and Latent GDF8P11M / no BMP1). Black labels are samples prior to the pull-down and arrows are following the aptamer pull-down. In (B), BMP1 processing of the latent complexes was confirmed via the detection of the ~20 kDa band (arrows). Non-pulled-down samples are shown in black, and the samples are pulled-down by aptamers below the “pull-down samples” labeling. Immunoblotting of latent complex is first used for anti-prodomain blotting, then stripped and reblotted for anti-mature ligand. All aptamer pull-downs were processed simultaneously. Experiments were independently repeated twice with similar results. D Schema of BMP1 processing required for aptamer binding. Source data are provided as a Source Data File.GDF11/8 ligand (monomer): 12.4 kDa, GDF11/8 ligand (dimer): 25 kDa, BMP1 processed prodomain: ~20 kDa, latent GDF11/8: ~40 kDa, mixture of unprocessed and latent GDF11/8: ~50 kDa. BMP1 = bone morphogenetic protein 1; TLD = Tolloid protease; spec = specific. Experiments performed in duplicate.

[See PDF for image]

Fig. 3

GDF11-specific aptamer did not recognize GDF11/8 mature ligand in the presence of their inhibitors.

A, B Anti-GDF11/8 mature ligands western blot of GDF8 (A) and GDF11 (B) mature ligands with inhibitors. The input shows the proteins before the pull-down. The proteins after the pull-down are below the “pulled-down sample” labeling, with non-pulled-down proteins on the left. Mature ligands and inhibitors are mixed in 1:2.5 molar ratio respectively. C Summary of ligand states captured by the aptamers. Experiments were independently repeated twice with similar results. GDF11/8 ligand (monomer): 12.4 kDa, GDF11/8 ligand (dimer): 25 kDa, FS288 = follistatin 288; FSTL3 = follistatin-like 3; WFIKKN1 = WAP, Kazal, immunoglobulin, Kunitz and NTR domain-containing protein 1; WFIKKN2 = WAP, Kazal, immunoglobulin, Kunitz and NTR domain-containing protein 2. Experiments performed in duplicate. Source data are provided as a Source Data File.

Immunoblotting following aptamer pull-down showed that all three aptamers captured negligible amounts of latent GDF8P11M or latent GDF8 (Fig. 2B, C). We next tested whether activation of GDF11/8 by prodomain cleavage is required for aptamer binding. The latent complex was cleaved by Bone Morphogenetic Protein 1 (BMP1), one of the tolloid family enzymes that can process the prodomain to form an active BMP1-processed latent complex in which the mature ligands remain bound to the prodomains^6,12,19. The dual-specific and GDF8-specific aptamers readily detected the mature ligand following cleavage with BMP1, whereas the GDF11-specific aptamer only slightly captured the mature ligand following cleavage with BMP1 (Fig. 2C, indicated by arrows and summarized in Fig. 2D). BMP1 processing of the latent complexes was confirmed via the detection of the ~20 kDa band shown in Fig. 2B (indicated by arrows).

Next, we tested whether the dual-specific, the GDF8 specific, or the GDF11 specific aptamers recognized the mature ligand when in complex with known binding proteins; these complexes are thought to form in vivo with the mature ligand after cleavage and loss of the prodomain. Following pull-down with the dual-specific aptamer, both GDF11 and GDF8 were well-detected in complex with the binding proteins (Fig. 3). Purified GDF8 and latent GDF8 from purified GDF8 complexes were well-detected by the GDF8 specific aptamer (Fig. 3A), but only purified GDF11 and GDF11:FS288 were detected by the GDF11 specific aptamer (Fig. 3B). A summary of the ligand states detected by the various aptamers is shown in Fig. 3C.

Biolayer Interferometry confirmed aptamer binding

To provide confirmatory evidence that aptamer 2765-4 bound to the ligand following activation of the latent complex, we performed Biolayer Interferometry (BLI), an optical technique for measuring molecular interactions in real time. As expected, the dual-specific aptamer 2765-4 was unable to detect latent GDF8 (Fig. 4A). However, treatment of the latent GDF8 complex with BMP1 for 24 h (Fig. 4B) and 48 h (Fig. 4C) led to robust detection of the ligand as indicated by the positive binding profiles observed in Fig. 4B, C. We confirmed the specificity of this signal by performing a dose-response in the HEK293 CAGA-luciferase assay with latent GDF8 and BMP1-treated latent GDF8 (Fig. 4D). As expected, we observed a higher luciferase response from BMP1-treated latent GDF8 than from latent GDF8 alone. Together these results show that BMP1 activation of the prodomain latent GDF8 is required for the aptamer 2765-4 to bind to the ligand and is consistent with our pull-down experiments described above.

[See PDF for image]

Fig. 4

Aptamer recognition of GDF8 following BMP1 activation of latent GDF8.

A previously performed on latent GDF8. B latent GDF8 treated with BMP1 for 24 h; (C) latent GDF8 treated with BMP1 for 48 h. The aptamer was coupled to a streptavidin biosensor and dipped into a dilution series ranging from 400 nM to 3.13 nM of latent GDF8 or BMP1-treated latent GDF8. The binding of the ligand to the aptamer is shown in black and the 1:1 binding model fit is shown below the “pulled-down samples” labeling. D Ligand activity was assessed at 24 h post BMP1 digestion using the HEK293 CAGA luciferase assay. Data shown are from at least 3 technical replicates. Source data are provided as a Source Data File.BLI = Biolayer interferometry; BMP1 = bone morphogenetic protein 1.

We also confirmed that aptamer 2765-4 readily recognizes the mature ligands, WFIKKN2:GDF11 and WFIKKN1:GDF11 but not WFIKKN2 or WFIKKN1 alone (Fig. S2). We observed specific binding of aptamer 2765-4 to the mature ligands and the antagonist:ligand complexes and no binding of the antagonists alone. As expected, we did not see any significant non-specific binding when samples were subjected to a control (scrambled) aptamer, suggesting that the 2765-4 aptamer selectively interacts with epitopes unique to the mature ligand and ligand-bound complexes rather than binding to the antagonists alone or exhibiting nonspecific interactions (Fig. S2). Together, these data support that the ligand binding to 2765-4 is specific for the mature ligand, and that it does co-bind antagonist:ligand complexes.

Baseline characteristics of the meta-cohort of six cardiovascular studies

Table 1 shows the baseline characteristics of the six independent cohorts (and their study fractions) which were pooled to create a single meta-cohort of 11,609 participants, with an overall 4 year primary outcome event rate of 21.9%. There were 2540 outcome events: 972 deaths (38%), 622 hospitalizations for heart failure (24.5%), 601 myocardial infarctions (23.6%), and 345 strokes (13.6%). Within the meta-cohort, there were two primary prevention, three secondary prevention, and one diabetes study, with 4 year event rates ranging widely from 10.8 to 49.7% among the six studies.

Table 1. Baseline characteristics of the 6 individual cohorts (and their study fractions) and their meta-cohort

Covariate	Measure	HUNT3 secondary	ARIC secondary	ARIC primary elderly	BASEL VIII secondary	BASEL VIII primary	EXSCEL Placebo baseline	Meta-cohort with 4-year outcomes
Population	Purpose/duration	Studies or study fractions with 4 year outcomes included in meta-cohort						All 4 year studies merged dataset
	Study fraction and Morbidity	20% of total secondary population	80% of total visit 5 secondary population	100% of visit 5 primary population aged >65	100% secondary w/CAD symptoms	100% primary w/CAD symptoms	100% of placebo; type 2 diabetes	Multi morbidity
	Sample size	139	784	4078	2410	1675	2523	11,609
Composite CV Event	Event (%)	37 (26.6%) Death=14 CHF = 10 MI = 10 Stroke = 3	390 (49.7%) Death=72 CHF = 207 MI = 80, Stroke = 31	817 (20.0%) Death=350 CHF = 253 MI = 109 Stroke=105	627 (26.0%) Death=236 CHF = 77, MI = 218 Stroke = 96	181 (10.8%) Death=124 CHF = 18 MI = 11 Stroke = 28	488 (19.3%) Death=176 CHF = 57, MI = 173 Stroke = 82	2540 (21.9%) Death=972 CHF = 622 MI = 601 Stroke = 345
	No event (%)	102 (73.4%)	394 (50.3%)	3261 (80.0%)	1783 (74.0%)	1494 (89.2%)	2035 (80.7%)	9069 (78.1%)
Follow-up time (days)	Mean events (SD) Range	638 (432) 4−1675	910 (616) 7 – 2232	1136 (580) 4-2375	776 (647) 1 - 2958	863 (679) 7 - 2822	725 (477) 1-1950	907 (614) 1 - 2958
	Mean no events (SD) Range	1721 (175) 1407 - 2035	2019 (257) 238 – 2389	2042 (235) 134-2404	1421 (669) 361 - 3024	1392 (681) 267 - 3054	1423 (346) 1-2112	1674 (558) 1 - 3054
Age (years)	Mean (SD) Range	69.5 (10.8) 44−92	77.2 (5.3) 67−90	75.3 (5.1) 66-90	68.1 (10.6) 34 - 93	65.9 (11.8) 26 - 95	62.8 (9.4) 29 – 88	69.8 (10.1) 26 - 95
Sex	Male (%)	108 (77.7%)	498 (63.5%)	1581 (38.8%)	1891 (78.5%)	855 (51.0%)	1520 (60.2%)	6453 (55.6%)
	Female (%)	31 (22.3%)	286 (36.5%)	2497 (61.2%)	519 (21.5%)	820 (49.0%)	1003 (39.8%)	5156 (44.4%)
Race	White (%)	n/a	652 (83.2%)	3322 (81.5%)	n/a	n/a	2032 (80.5%)	10,230 (88.2%)
	Black (%)	n/a	132 (16.8%)	756 (18.5%)	n/a	n/a	44 (1.7%)	932 (8%)
	Asian (%)	n/a	0	0	n/a	n/a	176 (7.0%)	176 (1.5%)
	Latine (%)	n/a	0	0	n/a	n/a	260 (10.3%)	260 (2.2%)
	Other (%)	n/a	0	0	n/a	n/a	11 (0.4%)	11 (0.10%)

CV cardiovascular, CAD coronary artery disease, CHF congestive heart failure, MI myocardial infarction, n/a not applicable, HUNT3 Trøndelag Health Study, ARIC Atherosclerosis Risk in Communities, BASEL VIII Biochemical and Electrocardiographic Signatures in the Detection of Exercise-induced Myocardial Ischemia, EXSCEL Exenatide Study of Cardiovascular Event Lowering.

Baseline characteristics of the dementia study

A total of 4288 participants at Atherosclerosis Risk in Communities (ARIC) Study visit 5 (2011–2013) were included in the analytic sample of the dementia study (Table S1). They were aged 75 years (SD = 5 years); 57.9% were women and 18.2% Black. At ARIC study visit 5, 78.9% (n = 3383) of participants were cognitively healthy and 21.1% (n = 905) met criteria for mild cognitive impairment. In total, n = 641 participants progressed to dementia after the blood draw for proteomic assessment over a time horizon of eight years.

Associations of aptamer-measured quartiles with cardiovascular outcomes

GDF11/8 quartiles measured with the dual-specific aptamer 2765-4 were significantly and inversely associated with the primary composite outcome as well as with each of its 4 individual event types (death, heart failure, myocardial infarction, stroke), shown in Fig. 5A, with p-values for trend for each of the event types provided in the Figure legend. When participants in the highest quartile of GDF11/8 were compared with those in the lowest quartile (Q4/Q1), the hazard ratio (95% CI) for the primary composite outcome was 0.43 (0.38-0.48), p = 9.1e-63, for all-cause mortality 0.33 (0.27-0.4), p = 4.8e-40, for myocardial infarction 0.75 (0.6–0.94), p = 0.036, for heart failure 0.28 (0.22-0.36), p = 1e-32 and for stroke 0.67 (0.5–0.9), p = 0.037.

[See PDF for image]

Fig. 5

Incidence of the composite end-point and heart failure hospitalization, stroke, myocardial infarction, and death in the meta-cohort unadjusted.

A Stratified by quartile of active GDF11/8 [Q1 (0–2.70) n = 2,905, Q2 (2.71-2.76) n = 2904, Q3 (2.77-2.83) n = 2899, Q4 (2.84-3.53) n = 2,901]. P-Values for trend are p = 2.0e-60 for the composite endpoint, p = 2.5e-39 for death, p = 2.8e-32 for heart failure, p = 6.3e-03 for myocardial infarction, and p = 2.3e-02 for stroke. B Stratified by quartile of active GDF8 [Q1 (0–2.27) n = 2903, Q2 (2.28-2.32) n = 2,903, Q3 (2.33-2.37) n = 2,901, Q4 (2.38-4.17) n = 2,902]. P-Values for trend are p = 2.6e-32 for the composite endpoint, p = 1.3e-17 for death, p = 3.7e-26 for heart failure, p = 6.8e-01 for myocardial infarction, and p = 1.3e-02 for stroke. C Stratified by quartiles of active GDF11 [Q1 (0–1.86) n = 2,930, Q2 (1.87–1.91) n = 2,884, Q3 (1.92–2.00) n = 2894, Q4 (2.01–4.94) n = 2901]. There were no significant trend associations between active GDF11 quartiles with the composite endpoint of any of its individual components.Statistical analysis: Chi-squared test for trend in proportions (two-sided) was used to assess each event group distribution for each respective GDF quartile set. Data are presented as mean ± SEM. Sample size reflects independent biological replicates, each representing a unique patient from the meta-cohort. No technical replicates were used. The unit of study is the individual patient. Groups compared were stratified by quartiles of circulating ligand levels. No pooling of samples occurred. Source data are provided as a Source Data File.

GDF8 quartiles measured with the GDF8-specific aptamer were also significantly and negatively associated with cardiovascular disease event rates for the composite outcome and for each of its individual event types except for myocardial infarctions, shown in Fig. 5B, with p-values for trend shown for each of the event types provided in the Figure legend. The hazard ratio (95% CI) for the primary composite outcome was 0.55 (0.49-0.62), p = 8.8e-35, for all-cause mortality 0.51 (0.42–0.61), p = 1.2e-17, for myocardial infarction 0.93 (0.74-1.2), p = 0.89, for heart failure 0.37 (0.29-0.46), p = 4.7e-29, and for stroke 0.63 (0.46-0.86), p = 0.033. GDF11 quartiles measured with the GDF11-specific aptamer showed no associations with the composite outcome or any of its individual components, as shown in Fig. 5C.

Associations of GDF11/8, GDF8 and GDF11 as continuous variables with cardiovascular outcomes

Analysis of the associations of GDF11/8, GDF8, and GDF11 expressed as continuous variables scaled to standard deviations (SD) was confined to those individuals within the meta-cohort who had all the required demographic and clinical co-variates available for multivariable adjustment. These variables included age, sex, race, study, smoking history, diabetes status, treatment for hypertension, systolic blood pressure, total cholesterol, and high-density lipoprotein cholesterol. These 5,526 individuals represented 4 of the studies (or their fractions) within the larger meta-cohort, consisting of ARIC primary age >65 years (n = 4,027), ARIC secondary events (n = 768), Exenatide Study of Cardiovascular Event Lowering (EXSCEL) trial placebo arm (n = 592) and the Trøndelag Health Study (HUNT3) (n = 139). The characteristics of these individuals are shown in Table 2. Their overall 4 year event rate (24.7%) is similar to that of the larger meta-cohort (21.9%). In unadjusted analyses per SD, GDF11/8 was significantly associated with the primary composite outcome (HR (CI) = 0.736 (0.695–0.780), p < 0.001) as was GDF8 (HR (CI) = 0.910 (0.855–0.970), p = 0.004) but not GDF11. These associations for GDF11/8 and GDF8 remained significant after full multivariable adjustment, i.e., Model 2 in Table 3. Among individual events, GDF11/8 was significantly associated with the risk of all-cause deaths, heart failure, and myocardial infarction in unadjusted and fully adjusted analyses and with a trend for the risk of stroke (p < 0.1) (Table 3). GDF8 was significantly associated with all-cause deaths in unadjusted and fully adjusted analyses but not with the other individual events (Table 3). GDF11 was not associated with any of the individual events (Table 3).

Table 2. Baseline characteristics of the individuals within meta-cohort with available demographic and clinical variables

Covariate	Measure	Total	Endpoint = CVD Event	Endpoint = No Event (Censored)
Sample Size	Number	5526	1361 (24.6%)	4165 (75.4%)
Age (years)	Mean (SD)	69.8 (10.1)	76.7 (6.8)	73.6 (6.5)
	Median	71	77	73
	Range	39 - 92	45 - 90	39 - 92
Follow-up Time (Days)	Mean (SD)	1738 (582)	1014 (598)	1974 (325)
	Median	1939	1021	2027
	Range	1–2404	4–2375	1 – 2404
Sex	Male	2569 (46.5%)	735 (54.0%)	1834 (44.0%)
	Female	2957 (53.5%)	626 (46.9%)	2331 (56.0%)
Race	White	4535 (82.1%)	1090 (80.1%)	3445 (82.7%)
	Black	917 (16.6%)	257 (18.9%)	660 (15.9%)
	Asian	25 (0.5%)	5 (0.4%)	20 (0.5%)
	Latine	46 (0.8%)	7 (0.5%)	39 (0.9%)
	Other / Unknown	3 (0.1%)	2 (0.1%)	1 (0.02%)
Smoking	Current	358 (6.5%)	90 (6.6%)	268 (6.4%)
	Former	2677 (48.4%)	707 (52.0%)	1970 (47.3%)
	Never	2092 (37.9%)	433 (31.8%)	1659 (39.8%)
	Unknown	399 (7.2%)	131 (9.6%)	268 (6.4%)
Diabetes	Yes	2144 (38.8%)	647 (47.6%)	1497 (35.9%)
	No	3382 (61.2%)	714 (52.4%)	2668 (64.1%)
Hypertension	Yes	3659 (66.2%)	1001 (73.6%)	2658 (63.8%)
	No	1867 (33.8%)	360 (26.4%)	1505 (36.2%)
HDL cholesterol (mg/dL)	Mean (SD)	50.9 (14.3)	48.9 (14.0)	51.6 (14.3)
	Median	49	47	50
	Range	16 - 132	16 - 128	16 -132
Total Cholesterol (mg/dL)	Mean (SD)	178 (42.2)	172 (43.5)	180 (41.6)
	Median	175	166	177
	Range	43 - 483	77 - 459	43 - 483
Systolic BP (mmHg)	Mean (SD)	130.2 (17.9)	131.6 (19.0)	129.7 (17.5)
	Median	129.0	130.5	128.5
	Range	66 - 229	67 - 205	66 – 229

CVD cardiovascular disease, HDL high-density lipoprotein, SD standard deviation, BP blood pressure.

Table 3. Association of GDF11/8, GDF8 and GDF11 analyzed as continuous variables with cardiovascular outcomes and mortality

Outcome	Model	Per 1 SD increase
		GDF11/8		GDF8		GDF11
		HR (95% CI)	p-value	HR (95% CI)	p-value	HR (95% CI)	p-value
Composite	Unadjusted (Model 0)	0.736 (0.695-0.780)	p < 0.001	0.910 (0.855-0.970)	p = 0.004	0.986 (0.934-1.041)	p = 0.613
	Model 1	0.709 (0.668-0.753)	p < 0.001	0.922 (0.867-0.980)	p = 0.009	0.980 (0.930-1.033)	p = 0.449
	Model 2	0.735 (0.692-0.781)	p < 0.001	0.933 (0.878-0.992)	p = 0.026	0.979 (0.929-1.032)	p = 0.438
Death	Unadjusted (Model 0)	0.711 (0.644-0.785)	p < 0.001	0.819 (0.722-0.928)	p = 0.002	1.048 (0.964-1.139)	p = 0.268
	Model 1	0.693 (0.624-0.769)	p < 0.001	0.856 (0.762-0.962)	p = 0.009	1.039 (0.959-1.125)	p = 0.348
	Model 2	0.717 (0.645-0.797)	p < 0.001	0.866 (0.771-0.973)	p = 0.016	1.037 (0.958-1.122)	p = 0.372
HF	Unadjusted (Model 0)	0.673 (0.611-0.742)	p < 0.001	0.931 (0.841-1.031)	p = 0.172	0.909 (0.817-1.010)	p = 0.077
	Model 1	0.653 (0.590-0.722)	p < 0.001	0.951 (0.864-1.047)	p = 0.307	0.901 (0.812-1.000)	p = 0.051
	Model 2	0.675 (0.609-0.748)	p < 0.001	0.966 (0.879-1.062)	p = 0.480	0.907 (0.817-1.006)	p = 0.064
MI	Unadjusted (Model 0)	0.833 (0.733-0.946)	p = 0.005	1.017 (0.904-1.144)	p = 0.775	1.010 (0.898-1.137)	p = 0.863
	Model 1	0.785 (0.686-0.897)	p < 0.001	0.990 (0.871-1.125)	p = 0.872	1.006 (0.900-1.123)	p = 0.920
	Model 2	0.812 (0.709-0.929)	p = 0.002	1.000 (0.881-1.135)	p = 1.000	1.002 (0.896-1.122)	p = 0.966
Stroke	Unadjusted (Model 0)	0.861 (0.731-1.014)	p = 0.073	0.881 (0.725-1.070)	p = 0.200	0.958 (0.809-1.135)	p = 0.622
	Model 1	0.826 (0.693-0.983)	p = 0.032	0.881 (0.726-1.069)	p = 0.200	0.958 (0.813-1.128)	p = 0.606
	Model 2	0.860 (0.720-1.026)	p = 0.093	0.885 (0.725-1.080)	p = 0.229	0.946 (0.802-1.116)	p = 0.512

Associations were assessed using Cox proportional hazards regression models (two-sided). GDF11/8, GDF8 and GDF11 levels were analyzed per standard deviation (SD) increase. Results are reported as hazard ratios (HRs) with 95% confidence intervals (CIs) and exact p-values. Bolded values indicate statistical significance (p  <  0.05). Model 1 adjusts for age, sex, race, and study. Model 2 further adjusts for smoking history, diabetes status, hypertension treatment, systolic blood pressure, total cholesterol, and HDL cholesterol.All models satisfied the proportional hazards assumption which was verified by Schoenfeld residuals. HF heart failure.

Consistency of results across the six different studies

The meta-cohort consisted of a total of 6 primary cardiovascular, secondary cardiovascular, and diabetes populations with the 4 year risk for the composite outcome ranging from a low 4 year risk of 10.8% (Basel VIII primary) to a high 4 year risk of 49.7% (ARIC secondary) (Table 1). Despite these and other differences in these populations, the relationships between GDF11/8, GDF8, and GDF11 with the primary outcome were consistent across all six studies (Fig. S3).

GDF11/8, GDF8, GDF11 and age

Figure 6 displays the relationship between GDF11/8, GDF8, and GDF11 levels and age in the meta-cohort. There was a significant graded relationship of GDF11/8 and GDF8 levels with age, with lower levels observed in the older individuals. GDF11 levels did not show a significant relationship to age.

[See PDF for image]

Fig. 6

Active GDF11/8, GDF8, and GDF11 ligand levels by age.

Box plots represent the distribution of circulating ligand levels measured by aptamer-based assays. The center line indicates the median, the bounds of the box represent the 25th–75th percentiles, and the whiskers extend to the 5th–95th percentiles. Units on the vertical axis are log10 transformed relative fluorescence units (RFU). Statistical analysis was performed using the Jonckheere–Terpstra test for trend (two-sided). A significant inverse relationship with age was observed for active GDF11/8 (p = 5.63e-130) and active GDF8 (p = 5.87e-253), while no significant trend was observed for active GDF11 (p = 0.331). Subjects per age group (years old): <50 (n = 497), 50-54 (n = 571), 55-59 (n = 830), 60−64 (n = 1,111), 65−69 (n = 1708), 70−74 (n = 2830), 75−79 (n = 2232), 80+ (1,830). Each data point represents an independent biological replicate corresponding to a unique individual. No technical replicates or pooled samples were used. The unit of study is the individual participant. Source data are provided as a Source Data File.

GDF11/8 as a potential biomarker of cardiovascular risk

While the focus of this study is on the cardiovascular pathobiology of GDF11/8 in humans and not on GDF11/8 as a potential biomarker of cardiovascular risk, for interested readers we have included predictive performance metrics for GDF11/8 by itself and when added to cardiovascular risk prediction model based on traditional risk factors (Fig. S4 and Table S2). In brief, based on the area under the curve (AUC) of the receiver operating characteristics (ROC) curves, GDF11/8 demonstrates similar or superior AUC for cardiovascular events as total cholesterol, high-density lipoprotein (HDL)-cholesterol, and systolic blood pressure, well-accepted clinical risk factors (Fig. S4).

Associations of GDF 11/8, GDF8 and GDF11 with dementia

Table 4 shows the associations of GDF11/8, GDF8, and GDF11 with the risk of incident dementia. GDF11/8 was significantly associated with the risk of dementia unadjusted (HR (CI) = 0.751 (0.597–0.945, p = 0.0147)) and fully adjusted for age, race, center, sex, education, APOEε4, eGFR-creatinine, baseline body mass index, diabetes status, treatment for hypertension, and smoking status (HR (CI) = 0.663 (0.514–0.854), p = 0.00148). GDF8 and GDF11 showed no significant association with the risk of dementia.

Table 4. Association of GDF11/8, GDF8 and GDF11 with Dementia

Outcome	Model	Per log2 increase
		GDF11/8		GDF8		GDF11
		HR (95% CI)	p-value	HR (95% CI)	p-value	HR (95% CI)	p-value
Dementia	Unadjusted (Model 0)	0.751 (0.597 – 0.945)	p = 0.0147	0.892 (0.715 – 1.114)	p = 0.313	0.999 (0.910 – 1.099)	p = 0.991
	Model 1	0.635 (0.495 – 0.814)	p = 0.00034	0.980 (0.796 – 1.207)	p = 0.849	0.988 (0.902 – 1.082)	p = 0.797
	Model 2	0.655 (0.511 – 0.839)	p = 0.00083	0.986 (0.801 – 1.213)	p = 0.892	0.993 (0.907 – 1.088)	p = 0.884
	Model 3	0.663 (0.514 – 0.854)	p = 0.00148	0.990 (0.805 – 1.217)	p = 0.925	0.995 (0.909 – 1.089)	p = 0.907

Associations were assessed using Cox proportional hazards regression models (two-sided) over an 8 year follow-up. GDF11/8, GDF8, and GDF11 were analyzed per log₂ increase in circulating ligand level. Results are reported as hazard ratios (HRs) with 95% confidence intervals (CIs) and exact p-values. Bolded values indicate statistical significance (p  <  0.05).

Model 1 adjusts for age, sex, race-center, education, and APOEε4 status.

Model 2 additionally adjusts for estimated glomerular filtration rate (eGFR-creatinine).

Model 3 additionally adjusts for body mass index, diabetes status, hypertension treatment, and smoking status.

All models met the proportional hazards assumption, which was verified by Schoenfeld residuals.

Discussion

The overarching purpose of this study was to explain the reported conflicting associations between circulating GDF8 and GDF11 levels and the risk of cardiovascular or incident dementia outcomes when these GDFs are quantified by an aptamer binding assay in contrast to tandem mass spectrometry (LC-MS/MS) assay. Higher levels detected using a dual-specific aptamer that recognizes both GDF11 and GDF8 (GDF11/8) predict lower risk of adverse cardiovascular and dementia outcomes, whereas total GDF11 and GDF8 levels measured by LC-MS/MS do not. Our biochemical studies revealed that the dual-specific aptamer GDF11/8 does not recognize the latent (immature) complex of either GDF11 or GDF8, but does recognize more mature forms of these ligands, including their activated and inhibitor-bound states. This biochemical discovery suggesting that the dual-specific GDF11/8 aptamer detects activated but not latent GDF11 and 8 led us to use this aptamer in several large, observational cohorts, representing a broad range of cardiovascular risk, showing that high levels of GDF11/8 are associated with a reduced risk of adverse cardiovascular outcomes and all-cause mortality. We also showed that higher levels of GDF11/8 are associated with lower risk of incident dementia. Lastly, circulating levels of GDF11/8 were lower in older individuals, raising the hypothesis that increased cardiovascular, mortality, and incident dementia risk with aging could be, at least in part, attributed to lower GDF11/8 levels.

Our biochemical findings highlight the role of tolloid proteases in GDF biology and potentially in human cardiovascular diseases. Tolloids (TLD) are zinc-dependent metalloproteinases that include four proteins: BMP1, mammalian tolloid (mTLD), tolloid-like 1 (TLL1), and TLL2. BMP1 and mTLD are splicing variants of the Bmp1 gene²⁰. TLD substrates are wide-ranging, and TLDs play important roles in the activation of TGFβ family ligands²¹, including GDF11 and GDF8. The prodomain of GDF8 is cleaved by all four members of the TLD family (preferentially by TLL2), whereas the prodomain of GDF11 is cleaved by only BMP1 and TLL1¹³. It is plausible that local activation of GDF11 by TLDs is an important step in linking GDFs to the risk of cardiovascular disease. It is also possible that other substrates of TLDs contribute to the observed outcomes, and while the dual-specific aptamer 2765-4 provides a robust biomarker for GDF11/8 activity, it does not directly demonstrate causal involvement in disease pathogenesis. Given the strength of the clinical outcome data, it is worth investigating the molecular mechanisms behind these findings.

Studies of the effects of GDF8 and GDF11 on cardiovascular and other phenotypes in various murine models have yielded conflicting results, dependent on experimental conditions, types of measurements, and murine models used^{22, 23–24}. When different murine models yield conflicting results, human studies become of paramount importance, particularly when they can be designed and conducted with scientific rigor. A prior study in humans by Olson, Ganz, and colleagues investigated the relationship of GDF11 and GDF8 to cardiovascular risk using the dual specific GDF11/8 aptamer 2765-4⁹. They studied 1899 patients with stable coronary heart disease from two cohorts (Heart and Soul cohort, based in the San Francisco Bay Area, and HUNT3 cohort, based in Norway). They found an inverse association between GDF11/8 and cardiovascular risk, with higher baseline levels of GDF11/8 associated with a lower risk of myocardial infarction, stroke, heart failure, and all-cause mortality over a follow-up of 8.9 years. An important limitation of that study was its narrow population, and thus uncertain generalizability, consisting solely of individuals with stable ischemic heart disease. The present study confirmed the generalizability of these findings by investigating more diverse populations.

Accordingly, our present clinical study was conducted in a meta-cohort totaling 11,609 participants. The meta-cohort was designed to be representative of the general population, consisting of primary and secondary cardiovascular risk cohorts as well as a cohort of individuals with type 2 diabetes⁵. We found a strong, inverse association of plasma levels of GDF11/8 with cardiovascular and mortality outcomes in each of these different clinical settings. There was a high consistency of findings in the 6 cohorts studied, which represent a broad range of risks. With the use of selective aptamers, findings with GDF11/8 correlated with GDF8 but not with GDF11. The effect size was larger for GDF11/8 than for GDF8, possibly reflecting different binding epitopes for these aptamers. We chose a population approach instead of meta-analysis as we had access to full relevant clinical data for each participant. It is important to point out that epidemiological associations in observational cohorts cannot prove causality, and thus the associations reported here do not prove that activated GDF8 or GDF11 are responsible for the outcomes. It is possible, for example, that activated GDF11 or GDF8 in the blood could be reporting increased tolloid enzyme activity in a specific location and that substrates other than GDF11 and GDF8 are important. It is important to note that we cannot exclude the possibility that the aptamers also recognize some other proteins.

Concerning incident dementia, Newman et al. measured GDF11 and GDF8 using an LC-MS/MS assay⁸. They found no relationship of GDF11 or GDF8 to incident dementia in the Cardiovascular Health Study (n = 1506) but found a direct (positive) relationship between GDF8 and the risk of dementia in the Health ABC Study (n = 1237). In contrast, we found a negative relationship between the levels of GDF11/8 measured with dual-specific aptamer 2765-4 and the risk of incident dementia in the ARIC cohort, suggesting that post-activation GDF11/8 may also serve as a marker of dementia risk and further highlight the link between cardiovascular and neurodegenerative diseases. Although the dementia etiology was not defined in the sample of participants used here, the large proportion of dementia cases was likely attributable to cerebrovascular disease, given the high prevalence of vascular cognitive impairment and dementia in the community. While the association of GDF11/8 with the risk of dementia was significant in the ARIC study, a large population with a sizeable number of dementia outcomes, confirmation of our findings in other cohorts will be important.

Our results offer a means by which the conflicting results between clinical studies that used LC-MS/MS versus aptamers can be potentially reconciled. Tandem mass spectrometry (MS/MS) is a time-honored method that digests proteins into peptides, each of which, when identified, serves as a surrogate for the protein molecule from which it is derived²⁵. The LC-MS/MS method reports the total amount of circulating mature and immature forms of GDF11 or GDF8 as these mature and immature forms share the peptides used to identify these proteins. Aptamers, however, recognize specific 3D epitopes of ligands in their native state, and thus can discriminate different ligand subforms and activity states^26,27. Our data suggest that the epitope that the dual-specific GDF11/8 aptamer measures forms primarily after the prodomain has been cleaved. This aptamer measurement provides not only a unique predictive measure from human blood but potentially a molecular mechanism in cardiovascular disease.

While the focus of this study is primarily on the biology that explains the association of GDFs with cardiovascular, mortality, and incident dementia outcomes, for interested readers we have included performance metrics for GDF11/8 as a biomarker of cardiovascular risk. Its prediction is similar or superior to well-accepted clinical risk factors such as HDL-cholesterol, total cholesterol, or systolic blood pressure. In our recent investigations of cardiovascular risk modeling, we recognize that a multi-biomarker approach is superior to any one individual biomarker. It is of interest regarding GDF11/8 that in our study of 1130 blood proteins, agnostic machine learning selected 9 proteins into the cardiovascular risk model, and among them was GDF11/8²⁸. In another study of ~5000 proteins, machine learning selected 27 proteins into the cardiovascular risk model, and among them again was GDF11/8⁵. Thus, machine learning algorithms have detected an important contribution that GDF11/8 makes to cardiovascular risk prediction. Our study does not demonstrate that the dual-specific aptamer measurement can be used to change clinical outcomes. Because the meta-cohort includes a broad population, we did not compare the dual-aptamer measure against other prediction models as these generally focus on a more focused population.

Clinical Implications: These data explain why some assays (e.g. LC-MS/MS) that measure total GDF11/8 in human blood may not predict clinical outcome, whereas high levels of subforms of GDF11/8, measured by dual-specific aptamer technology, are highly predictive of favorable cardiovascular, dementia, and all-cause mortality outcomes in a broad population of patients with primary prevention, secondary prevention, or type 2 diabetes. These data also point to the activation of GDF11/8 latent complexes as a potential key step in the natural history of cardiovascular disease, all-cause mortality, and dementia. Future studies should elucidate whether GDFs should be explored as a potential therapeutic target to improve outcomes.

Methods

Ethics statement

All research involving human participants was conducted in accordance with relevant ethical regulations. Study protocols for each cohort were approved by the local institutional review boards or ethics committees at their respective coordinating centers: the University of North Carolina at Chapel Hill (ARIC), the Norwegian University of Science and Technology (HUNT3), the University of Basel (BASEL VIII), and the Duke Clinical Research Institute (EXSCEL). All participants provided written informed consent.

Patient populations for cardiovascular and mortality outcomes

The study of the associations of GDF11/8, GDF8, and GDF11 with cardiovascular and mortality outcomes was conducted in six separate cohorts (described below) that were pooled into a single ‘meta-cohort’ consisting of a total of 11,609 participants⁵. This meta-cohort was previously analyzed to validate a 27-protein cardiovascular risk model⁵. The six cohorts within the meta-cohort represent primary and secondary cardiovascular event populations with a broad representation of epidemiologically observed risks⁵: (1) random 20% of the secondary event population in The Trøndelag Health Study (HUNT3), (2) random 80% of the secondary event population at visit 5 in the Atherosclerosis Risk in Communities (ARIC) Study, (3) 100% of visit 5 ARIC primary event population aged >65 years, (4) 100% of the secondary event population with coronary artery disease symptoms from the Biochemical and Electrocardiographic Signatures in the Detection of Exercise-induced Myocardial Ischemia (BASEL VIII) study, (5) 100% of the primary event population of BASEL VIII, and (6) placebo arm from the Exenatide Study of Cardiovascular Event Lowering (EXSCEL) Trial. Each of these cohorts is described in detail in the Supplement.

Patient populations for dementia outcomes

The study of the associations of GDF11/8, GDF8, and GDF11 with incidence of dementia was conducted in the ARIC Study, with blood samples analyzed from the fifth study visit. ARIC is a prospective community-based study of cardiovascular disease and its risk factors²⁹. At baseline (1987−89), 15,792 men and women aged 45−64 years were recruited from 4 communities in the US (Washington County, MD; Forsyth County, NC; Jackson, MS; Minneapolis, MN). The fifth visit was conducted in 2011–2013, and follow-up was to visit 7 conducted in 2018–2019. Time horizon for this analysis was 8 years.

Cardiovascular and mortality study outcomes

The primary outcome in this study was defined as the first event subsequent to the blood sample, among myocardial infarction (MI), stroke, heart failure hospitalization, or all-cause death, over a 4 year time horizon. This composite endpoint employed in our prior proteomics studies^5,9,28,30,31 captures a broad range of cardiovascular events and all-cause-death. As previously^5,28,30,31, we chose all-cause death because of mounting evidence that clinical adjudication of cardiovascular from non-cardiovascular causes tends to be inaccurate³², and because the risk of this more encompassing outcome may be of greatest interest to patients and health care providers. For brevity, we refer to this compositive endpoint as ‘cardiovascular outcome’. The four individual components of the primary composite outcome constituted secondary outcomes.

Incidence of dementia

ARIC Dementia Classification. The ARIC dementia assessment has been described in detail elsewhere³³. Briefly, at ARIC visits 5 (2011- 13), 6 (2016-17), and 7 (2018–19), dementia was classified based on a comprehensive cognitive and functional evaluation, using criteria defined by the National Institute on Aging-Alzheimer’s Association (NIA-AA) workgroups³⁴ and the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-V)³⁵. Dementia was defined using an algorithm followed by adjudication by a panel of experts. Between visits, participants’ cognitive status was identified using the Ascertain Dementia 8-item Questionnaire (AD8)³⁶ and the Six-Item Screener (SIS)³⁶ administered annually via phone. Dementia surveillance also included an examination of the International Classification of Diseases, Ninth Revision (ICD-9) hospital discharge codes, and diagnostic codes from death certificates. For participants who were determined as having dementia at visits 6 and 7 via cognitive and functional assessment, the AD8, SIS, and ICD-9 hospital discharge codes were used to define date of dementia onset. For participants who attended visit 5 but did not attend visits 6 and/or 7, dementia diagnosis and date of dementia onset were defined using the AD8, SIS, ICD-9 hospital discharge codes, and diagnostic codes from death certificates.

Measurements of GDF11/8, GDF8, and GDF11

The plasma levels of GDF11/8, GDF8, and GDF11 were measured by a modified aptamer binding assay using the SomaScan^® v.4 platform^{28,30,37, 38–39}. GDF11/8 represents the sum of GDF11 and GDF8 as it is measured with a dual-specific aptamer that has equal affinity for the two nearly identical proteins. GDF11/8 has been previously reported to have a strong inverse association with cardiovascular and mortality outcomes in patients with stable coronary heart disease⁹. Since that initial report, selective aptamers have been developed for GDF8 and GDF11 and employed here, and the SELEX procedure has been previously described in detail¹⁰. Characteristics of the GDF11/8, GDF8, and GDF11 aptamers, including their precision (coefficients of variation), affinity, and cross-reactivity, are presented in Supplemental Table 3. The details of the SomaScan binding assay and its performance characteristics are described in the Supplement. The aptamer sequences are:

Aptamer 2765-4 (GDF11/8):

5’-gggtcAZCGGAZAZGAZZAZGAAAGGGGGAZZGZZZGGGAGZAAgaccc-3’

Aptamer 12060-28 (GDF11):

5′-ccctgCGCCPPCGGACPPGCPPPAAGPPPAGCCGCPPGCPCACAPcacaa-3′

Aptamer ID 14583-49 (GDF8):

5’ctctgAWACGGCCWWAGWGWWAGAGWCCWWWAWGAGWCWAGACCAcaaca-3’

Letters Z, P, and W in the sequences indicate 5-[N-benzylcarboxamide]-2’-deoxyuridine (Z, BndU), 5-[N-(1-naphthylmethyl)carboxamide]-2’-deoxyuridine (P, NapdU) or 5-[N-(tryptamino)carboxamide]-2’-deoxyuridine (W, TrpdU); lowercase and uppercase letters indicate nucleotides in the originally fixed-sequence and randomized regions, respectively.

Statistical analysis

Cardiovascular and mortality analyses

We divided participants into quartiles by plasma GDF11/8, GDF8, and GDF11 levels. We calculated event rates for each respective GDF quartile and plotted the event rates with 95% CI from 100 bootstrap resamples. We assessed each event group distribution for each respective GDF quartile set using chi-squared test for trend in proportions. Additionally, plasma GDF distributions were tested for trends against 5 year age intervals using the Jonckheere–Terpstra trend test.

We also evaluated the association between cardiovascular and mortality events and levels of GDF11/8, GDF8, and GDF11 as a continuous variable. Proportional hazard assumptions and the absence of co-linearity were verified for all models. No evidence of departure from linearity was present in the continuous GDF11/8, GDF8, and GDF11 models. Continuous models are presented unadjusted and adjusted for age, sex, race, total cholesterol, high-density lipoprotein (HDL)-cholesterol, smoking status, diabetes status, treatment for hypertension, and systolic blood pressure, and study as appropriate. The adjustment covariates were selected a priori, based on the literature. In the publication of the cohorts that constitute the meta-cohort, patients were described as Male or Female, based on self-report, without additional explanation.

Analyses were performed using R version 4.1.0. All reported P-values are two-sided, with a P-value of 0.05 considered to indicate statistical significance.

Dementia analyses

We used Cox proportional hazard regression models to examine the association of GDF11/8, GDF8, and GDF11 with 8 year incident dementia risk in the ARIC study. Levels of GDF11/8, GDF8, and GDF11 were measured at ARIC visit 5 (2011–13). Participants with dementia at baseline (visit 5) were excluded from analyses. New-onset dementia occurring between ARIC visit 5 and 7 (2018–19) was considered in these time-to-event analyses. We examined an unadjusted model (model 0) and 3 covariate-adjusted models. Model 1 adjusted for demographic variables (age, sex, race-center, education) and APOEε4 status. Model 2 additionally adjusted for kidney function, defined using eGFR-creatinine. Model 3 additionally adjusted for cardiovascular risk factors (BMI, diabetes status, treatment for hypertension, and smoking status). The adjustment covariates were selected a priori, based on the literature. The Cox proportionality assumption was tested by computing and plotting Schoenfeld residuals. Covariates that did not meet the proportional hazards assumption were incorporated in sensitivity analyses as stratified variables or with a time interaction (covariate time) to determine whether results differed from those derived in the primary analyses. Analyses were conducted using R v3.6.2.

Ethics and inclusion statement

The research presented here included local and national collaborators. The research is not locally relevant. All roles and responsibilities were agreed amongst all collaborators and authors. This research would not have been strictly restricted in the setting of the researchers. All experimental data is available upon request. All human data is available upon request, consistent with the policies of each cohort represented in the study.

Reporting summary

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

Acknowledgements

We gratefully acknowledge the efforts of the steering committees and participants for the use of the SomaScan and outcomes data in the following studies: ARIC, HUNT3, BASEL VIII, and EXSCEL. We thank SomaLogic, Inc. for providing research support and aptamers to the Lee Lab. This work was supported by NIH grants R56AG062468, R01AG047131, HL169291 (to RT Lee), R01HL059367 and R01HL153499-02 (to Peter Ganz), 4T32HL007208-39 (to RG Walker), R35GM134923 and R01AG072087 (to TB Thompson), R01AG057428 and an award from the Glenn Foundation (to AJ Wagers), R01AG072086 (to L Rubin and AJ Wagers), and 1R56AG052972 and P30AG31679 to SB. KA Walker received funding from the NIA Intramural Research Program. JE Walter is supported by Swiss Heart Foundation grants FF19097 and F18111. This study was funded, in part, by the NIA Intramural Research Program.

Author contributions

R.G.W., T.K, L.B.D., N.J., A.D.G., K.R.M., M.A.C., J.V.G., L.L.R., A.J.W., T.B.T., P.G., and R.T.L. contributed to the Design, the experiments, the analysis of the data, and the writing of the manuscript. S.A.W., M.A.H., C.K., J.E.W., C.M., K.A.W., J.C., and S.B. contributed to the acquisition and analysis of the clinical data and the preparation and editing of the manuscript.

Peer review

Peer review information

Nature Communications thanks Yingfu Li, and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The clinical and proteomic data used in this study were generated by the ARIC, HUNT3, EXSCEL, and BASEL VIII cohort studies and are available under restricted access due to participant confidentiality policies.

Access can be obtained by submitting a request to the respective study steering committees:

• ARIC (Atherosclerosis Risk in Communities Study): https://sites.cscc.unc.edu/aric/

• HUNT3 (The Trøndelag Health Study): https://www.ntnu.edu/hunt/data

• EXSCEL (Exenatide Study of Cardiovascular Event Lowering Trial): https://clinicaltrials.gov/ct2/show/NCT01144338

• BASEL VIII: https://clinicaltrials.gov/ct2/show/NCT01838148

The raw clinical and proteomic data are protected and are not publicly available due to data privacy laws. The processed data used to generate the figures and tables in this study will be made available upon reasonable request from the corresponding author within 2 weeks and can be freely used. No custom code was used beyond standard statistical packages in R. All other data supporting the findings of this study are available within the article and its Supplementary Information files.  are provided with this paper.

Competing interests

R.T.L., R.G.W., L.L.R., A.J.W, and T.B.T. are inventors of patents related to GDF11 and GDF8 through their institutions which are directly related to the proteins studied in this work. SomaLogic, Inc. provided research support and aptamers used in this study to the Lee Lab. J.E.W. was employed at Roche Diagnostics Switzerland in 2021 for 9 months as an expert in personalized healthcare oncology, unrelated to the present work. P.G. has served on the medical advisory board of SomaLogic without receiving financial compensation. S.B. has received research grants from AbbVie, MIB, and FPT, and consultation fees from Novartis for work unrelated to the present study. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-025-61815-w.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Loffredo, FS et al. Growth differentiation factor 11 is a circulating factor that reverses age-related cardiac hypertrophy. Cell; 2013; 153, pp. 828-839.1:CAS:528:DC%2BC3sXnsFWisrc%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23663781][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3677132]

2. Katsimpardi, L et al. Vascular and neurogenic rejuvenation of the aging mouse brain by young systemic factors. Science; 2014; 344, pp. 630-634.2014Sci..344.630K1:CAS:528:DC%2BC2cXnsVSqsr0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24797482][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4123747]

3. Sinha, M et al. Restoring systemic GDF11 levels reverses age-related dysfunction in mouse skeletal muscle. Science; 2014; 344, pp. 649-652.2014Sci..344.649S1:CAS:528:DC%2BC2cXnsVSqs7k%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24797481][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4104429]

4. Harper, SC et al. Is growth differentiation factor 11 a realistic therapeutic for aging-dependent muscle defects?. Circ. Res.; 2016; 118, pp. 1143-1150.1:CAS:528:DC%2BC28XltFGgtrw%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27034276][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4829942]

5. Williams, SA et al. A proteomic surrogate for cardiovascular outcomes that is sensitive to multiple mechanisms of change in risk. Sci. Transl. Med.; 2022; 14, 1:CAS:528:DC%2BB38XhtVSht7jM [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35385337]eabj9625.

6. Walker, RG et al. Biochemistry and biology of GDF11 and myostatin: similarities, differences, and questions for future investigation. Circ. Res.; 2016; 118, pp. 1125-1141.2016teas.book...W1:CAS:528:DC%2BC28XltFGgsbc%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27034275][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4818972]

7. Schafer, MJ et al. Quantification of GDF11 and myostatin in human aging and cardiovascular disease. Cell Metab.; 2016; 23, pp. 1207-1215.1:CAS:528:DC%2BC28Xptl2mur8%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27304512][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4913514]

8. Newman, AB et al. Evaluation of associations of growth differentiation factor-11, growth differentiation factor-8, and their binding proteins follistatin and follistatin-like protein-3 with dementia and cognition. J. Gerontol. A Biol. Sci. Med. Sci.; 2023; 78, pp. 2039-2047.1:CAS:528:DC%2BB2cXhsVCiurzM [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36660892][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10613013]

9. Olson, KA et al. Association of growth differentiation factor 11/8, putative anti-ageing factor, with cardiovascular outcomes and overall mortality in humans: analysis of the Heart and Soul and HUNT3 cohorts. Eur. Heart J.; 2015; 36, pp. 3426-3434.1:CAS:528:DC%2BC1cXpsFCisg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26294790][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4685178]

10. Ochsner, UA et al. Targeting unique epitopes on highly similar proteins GDF-11 and GDF-8 with modified DNA aptamers. Biochemistry; 2019; 58, pp. 4632-4640.1:CAS:528:DC%2BC1MXhvF2ku7jK [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31638376]

11. Lee, SJ; McPherron, AC. Regulation of myostatin activity and muscle growth. Proc. Natl. Acad. Sci. USA; 2001; 98, pp. 9306-9311.2001PNAS..98.9306L1:CAS:528:DC%2BD3MXlvFSrsL0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/11459935][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC55416]

12. Wolfman, NM et al. Activation of latent myostatin by the BMP-1/tolloid family of metalloproteinases. Proc. Natl. Acad. Sci. USA; 2003; 100, pp. 15842-15846.2003PNAS.10015842W1:CAS:528:DC%2BD2cXhtVCltg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/14671324][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC307655]

13. Ge, G; Hopkins, DR; Ho, W-B; Greenspan, DS. GDF11 forms a bone morphogenetic protein 1-activated latent complex that can modulate nerve growth factor-induced differentiation of PC12 cells. Mol. Cell. Biol.; 2005; 25, pp. 5846-5858.1:CAS:528:DC%2BD2MXmvVeit7o%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15988002][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1168807]

14. Walker, RG et al. Molecular characterization of latent GDF8 reveals mechanisms of activation. Proc. Natl. Acad. Sci. USA; 2018; 115, pp. E866-E875.1:CAS:528:DC%2BC1cXhtlKqtr0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29348202][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5798348]

15. Hill, JJ et al. The myostatin propeptide and the follistatin-related gene are inhibitory binding proteins of myostatin in normal serum. J. Biol. Chem.; 2002; 277, pp. 40735-40741.1:CAS:528:DC%2BD38XnvFyntb0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/12194980]

16. Hill, JJ; Qiu, Y; Hewick, RM; Wolfman, NM. Regulation of myostatin in vivo by growth and differentiation factor-associated serum protein-1: a novel protein with protease inhibitor and follistatin domains. Mol. Endocrinol.; 2003; 17, pp. 1144-1154.1:CAS:528:DC%2BD3sXktlymtrk%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/12595574]

17. Kondas, K; Szlama, G; Trexler, M; Patthy, L. Both WFIKKN1 and WFIKKN2 have high affinity for growth and differentiation factors 8 and 11. J. Biol. Chem.; 2008; 283, pp. 23677-23684.1:CAS:528:DC%2BD1cXhtVWktr%2FO [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18596030][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3259755]

18. Walker, RG et al. Structural basis for potency differences between GDF8 and GDF11. BMC Biol.; 2017; 15, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28257634][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5336696]19.

19. McFarlane, C et al. Proteolytic processing of myostatin is auto-regulated during myogenesis. Dev. Biol.; 2005; 283, pp. 58-69.1:CAS:528:DC%2BD2MXlsFGitLw%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15975431]

20. Vadon-Le Goff, S; Hulmes, DJS; Moali, C. BMP-1/tolloid-like proteinases synchronize matrix assembly with growth factor activation to promote morphogenesis and tissue remodeling. Matrix Biol.; 2015; 44-46, pp. 14-23.1:CAS:528:DC%2BC2MXjsVahtrs%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25701650]

21. Ge, G; Greenspan, DS. BMP1 controls TGFbeta1 activation via cleavage of latent TGFbeta-binding protein. J. Cell Biol.; 2006; 175, pp. 111-120.1:CAS:528:DC%2BD28XhtFSisLjK [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/17015622][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2064503]

22. Harper, SC et al. GDF11 decreases pressure overload-induced hypertrophy, but can cause severe cachexia and premature death. Circ. Res.; 2018; 123, pp. 1220-1231.2018frcd.book...H1:CAS:528:DC%2BC1cXitFWmurzN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30571461][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6309347]

23. Smith, SC et al. GDF11 does not rescue aging-related pathological hypertrophy. Circ. Res.; 2015; 117, pp. 926-932.1:CAS:528:DC%2BC2MXhslOgt77O [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26383970][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4636963]

24. Poggioli, T et al. Circulating growth differentiation factor 11/8 levels decline with age. Circ. Res.; 2015; 118, pp. 29-37. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26489925][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4748736]

25. Smith, LM; Kelleher, NL. Proteoforms as the next proteomics currency. Science; 2018; 359, pp. 1106-1107.2018Sci..359.1106S1:CAS:528:DC%2BC1cXlt1yqtb8%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29590032][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5944612]

26. Alexaki, A et al. Effects of codon optimization on coagulation factor IX translation and structure: Implications for protein and gene therapies. Sci. Rep.; 2019; 9, 2019NatSR..915449A [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31664102][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6820528]15449.

27. Jankowski, W et al. Modified aptamers as reagents to characterize recombinant human erythropoietin products. Sci. Rep.; 2020; 10, 2020NatSR.1018593J1:CAS:528:DC%2BB3cXit1Oqs77K [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33122796][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7596557]18593.

28. Ganz, P et al. Development and validation of a protein-based risk score for cardiovascular outcomes among patients with stable coronary heart disease. JAMA; 2016; 315, pp. 2532-2541.1:CAS:528:DC%2BC28XhtlOlsbzK [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27327800]

29. Wright, JD et al. The ARIC (Atherosclerosis risk in communities) study: JACC focus seminar 3/8. J. Am. Coll. Cardiol.; 2021; 77, pp. 2939-2959. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34112321][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8667593]

30. Williams, SA et al. Plasma protein patterns as comprehensive indicators of health. Nat. Med.; 2019; 25, pp. 1851-1857.1:CAS:528:DC%2BC1MXit1yjsrrI [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31792462][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6922049]

31. Williams, SA et al. Improving assessment of drug safety through proteomics: early detection and mechanistic characterization of the unforeseen harmful effects of torcetrapib. Circulation; 2018; 137, pp. 999-1010.1:CAS:528:DC%2BC1cXjvVOnsbw%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28974520][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5839936]

32. Tseng, ZH et al. Prospective countywide surveillance and autopsy characterization of sudden cardiac death: post scd study. Circulation; 2018; 137, pp. 2689-2700. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29915095][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6013842]

33. Walker, KA et al. Large-scale plasma proteomic analysis identifies proteins and pathways associated with dementia risk. Nat. Aging; 2021; 1, pp. 473-489. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37118015][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10154040]

34. McKhann, GM et al. The diagnosis of dementia due to Alzheimer’s disease: recommendations from the National Institute on Aging-Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimers Dement.; 2011; 7, pp. 263-269. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21514250]

35. Battle, DE. Diagnostic and statistical manual of mental disorders (DSM). Codas; 2013; 25, pp. 191-192. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24413388]

36. Carpenter, CR; DesPain, B; Keeling, TN; Shah, M; Rothenberger, M. The six-item screener and AD8 for the detection of cognitive impairment in geriatric emergency department patients. Ann. Emerg. Med; 2011; 57, pp. 653-661. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20855129]

37. Brody, E et al. Life’s simple measures: unlocking the proteome. J. Mol. Biol.; 2012; 422, pp. 595-606.1:CAS:528:DC%2BC38XpvVOkt7s%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22721953]

38. Rohloff, JC et al. Nucleic acid ligands with protein-like side chains: modified aptamers and their use as diagnostic and therapeutic agents. Mol. Ther. Nucleic Acids; 2014; 3, 1:CAS:528:DC%2BC2cXhslSrt7rN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25291143][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4217074]e201.

39. Kim, CH et al. Stability and reproducibility of proteomic profiles measured with an aptamer-based platform. Sci. Rep.; 2018; 8, 2018NatSR..8.8382K [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29849057][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5976624]8382.